Sādhanā Vol. 36, Part 5, October 2011, pp. 651–697.


c Indian Academy of Sciences

Epoch-based analysis of speech signals

B YEGNANARAYANA∗ and SURYAKANTH V GANGASHETTY

International Institute of Information Technology, Hyderabad 500 032, India

e-mail: [email protected]; [email protected]

Abstract. Speech analysis is traditionally performed using short-time analysis to

extract features in time and frequency domains. The window size for the analysis is
fixed somewhat arbitrarily, mainly to account for the time varying vocal tract system
during production. However, speech in its primary mode of excitation is produced due
to impulse-like excitation in each glottal cycle. Anchoring the speech analysis around
the glottal closure instants (epochs) yields significant benefits for speech analysis.
Epoch-based analysis of speech helps not only to segment the speech signals based
on speech production characteristics, but also helps in accurate analysis of speech.
It enables extraction of important acoustic-phonetic features such as glottal vibra-
tions, formants, instantaneous fundamental frequency, etc. Epoch sequence is useful
to manipulate prosody in speech synthesis applications. Accurate estimation of epochs
helps in characterizing voice quality features. Epoch extraction also helps in speech
enhancement and multispeaker separation. In this tutorial article, the importance of
epochs for speech analysis is discussed, and methods to extract the epoch informa-
tion are reviewed. Applications of epoch extraction for some speech applications are
demonstrated.

Keywords. Epoch; zero-frequency filtering; fundamental frequency; pitch;

impulse-like excitation.

1. Significance of epochs in speech analysis

Speech is the output of a time-varying vocal tract system excited by a time-varying excitation. In
the resulting speech signal, the information of the speech production system is embedded as rela-
tions in the sequence of values of the signal at different instants of sampling the signal. The main
objective of speech signal processing is to extract the information of the time varying character-
istics of the speech production system. The information is represented in the form of parameters
or features derived from the signal. Knowledge at different levels, such as acoustic-phonetic,
prosody, lexical, syntactic, etc. is used to interpret the message in the speech signal from the
sequence of parameter or feature vectors. Thus, an algorithmic way of extracting the informa-
tion in the speech signal involves operations of representation (interms of extracted parameters
or features) and processing (to extract the information or message), in that order.

∗
For correspondence

651
652 B Yegnanarayana and Suryakanth V Gangashetty

On the other hand, the information or message in the speech, extracted in direct listening
by human beings, can be explained in the operations of processing and representation, in that
order. This is called as speech processing by human beings as opposed to speech signal pro-
cessing by a machine. In speech processing, the input speech acoustic wave is processed first to
extract the relevant information such as phonemes, syllables, etc., and the extracted information
is represented in the form of sequence of symbols, similar to text as in dictation taking. Thus,
speech processing is an evolving human activity, besides the sophisticated auditory processing
and selective attention to decide which part is to be processed. It is likely that human process-
ing may not consider all of the acoustic wave input as equally important. The processing may
be interpreted in terms of events in the input at various levels, such as at the signal level, speech
production level, prosody level, language level, etc. These events may be derived from features
present at subsegmental (1–3 ms), segmental (10–30 ms) and suprasegmental (>300 ms) levels.
Methods for speech signal processing mostly use segmental analysis using a fixed
frame/window size in the range (10–30 ms). The methods include model-free short-time spec-
trum analysis or model-based linear prediction (LP) analysis over each (frame size) segment of
data (Makhoul 1975). Although the knowledge of speech production and perception is invoked in
some analysis techniques such as source-system model or auditory perception based filter-bank
type analysis, all the speech signal data is used in the initial processing for extracting parameters
or features to represent the information in the speech signal. Since human processing of speech
seems to be anchored around events, there is need to explore new methods of processing speech
signal which are event-based.
As mentioned, since events can occur at several levels, it is necessary to first identify such
events before further processing of the signal. In this paper, the events at the production level are
considered, and in particular, only the significant events occurring due to major source of excita-
tion (glottal vibration) at epochs are considered for further processing to extract information in
the speech signal.
Voiced speech analysis consists of determining the glottal pulses representing the excitation
source and frequency response of the vocal-tract system. Although, the source of excitation for
voiced speech is a sequence of pulses at the glottis, the significant excitation of the vocal-tract
system within a glottal pulse can be considered to occur at the instant of glottal closure (GCI),
called the epoch (Gauffin & Sundberg 1989). Many speech analysis situations depend on accu-
rate estimation of the location of the epoch within a glottal pulse. For example, knowledge of
the epoch locations is useful for accurate estimation of the fundamental frequency (F0 ). Often
the glottal airflow is zero soon after the glottal closure. As a result, the supralaryngeal vocal
tract is acoustically decoupled from the trachea. Hence, the speech signal in the closed glottis
region represents free resonances of the supralaryngeal vocal-tract system. Analysis of speech
signals in the closed glottis regions provides an accurate estimate of the frequency response
of the supralaryngeal vocal-tract system (Murty 2009; Veeneman & BeMent 1985). With the
knowledge of the epochs, it may be possible to determine the characteristics of the voice source
by careful analysis of the signal within a glottal pulse. The epochs can be used as pitch mark-
ers for prosody manipulation, which is useful in applications such as text-to-speech synthesis,
voice conversion and speech rate conversion (Hamon et al 1989; Rao & Yegnanarayana 2006;
Yegnanarayana & Veldhuis 1998). Knowledge of the epoch locations may be useful for esti-
mating the time-delay between speech signals collected over a pair of spatially separated
microphones (Yegnanarayana et al 2005). The segmental signal-to-noise ratio (SNR) of the
speech signal is high in the regions around the epochs, and hence it is possible to enhance the
speech by exploiting the characteristics of speech signals around the epochs (Yegnanarayana &
Murthy 2000). It has been shown that the excitation features derived from the regions around
 Epoch-based analysis of speech signals 653

the epoch locations provide complementary speaker-specific information to the existing spectral
features (Murty & Yegnanarayana 2006; Neocleous & Naylor 1998; Plumpe et al 1999).
As a result of significant excitation at the epochs, the regions in the speech signal that immedi-
ately follow the epochs are relatively more robust to (external) degradations than other regions.
The instants of significant excitation play an important role in human perception also. It is
because of the epochs in speech that human beings seem to be able to perceive speech even at a
distance (e.g., 10 feet or more) from the source, even though the direct signal suffer an attenua-
tion of over 40 dB. For example, we may not be able to get the message in whispered speech by
listening to it at a distance of even 10 feet due to absence of regular epochs. The neural mech-
anism of human beings seems to have the ability of processing selectively the robust regions
around the epochs for extracting the acoustic cues even under degraded conditions. It is the abil-
ity of human beings to focus on these microlevel events that may be responsible for perceiving
speech information even under severe degradation such as noise, reverberation, presence of other
speakers and channel variations.
This paper discusses the importance of epochs for speech analysis, reviews epochs extrac-
tion methods and illustrates a few applications of epoch-based analysis of speech. The paper is
organized as follows (see table 1). Section 2 gives a review of epoch extraction methods which
includes methods based on analysis of speech signal. In particular, epoch extraction method
based on zero-frequency filtering (ZFF) and the robustness of the method in comparison with
other methods are discussed in some detail. Section 3 deals with the main theme of the paper,

Table 1. Table of contents.

1. Significance of epochs in speech analysis

2. Review of epoch extraction methods
2.1 Epoch extraction from electroglottography
2.2 Epoch extraction from speech signals
2.3 DYPSA algorithm for epoch extraction
2.4 Epoch extraction using zero-frequency resonator
2.5 Comparison of epoch extraction methods
3. Epoch-based speech analysis
3.1 Glottal activity detection
3.2 Estimation of instantaneous fundamental frequency
3.3 Estimation of formant frequencies
3.4 Glottal activity in stop consonant analysis
3.5 Determination of phonetic features of glottalization, creaky voice and glottal stop
3.6 Characterization of loudness of speech
3.7 Analysis of Lombard effect speech
3.8 Analysis of laugh signals
3.9 Analysis of speech produced at different speaking rates
3.10 Processing of speech collected at a distance
3.11 Pitch extraction from multispeaker data
4. Application of epoch-based analysis
4.1 Time delay estimation and determining number of speakers from mixed signals
4.2 Speech enhancement in single channel case
4.3 Multichannel speech enhancement
4.4 Multispeaker separation
4.5 Prosody modification using knowledge of epoch location
5. Practical issues in developing speech system using epoch-based analysis
654 B Yegnanarayana and Suryakanth V Gangashetty

namely, how to exploit the epoch location information for speech analysis. It is shown that
knowledge of epochs can be used for analysis of speech for glottal activity detection, extrac-
tion of instantaneous fundamental frequency, deriving voice quality features from speech such
as loudness and characterization of laughter, creaky voice, etc. Some applications are illustrated
in section 4, which include time delay estimation, speech enhancement from single and multi-
channel data and separation of multispeaker data from mixed signals. Section 5 discusses some
practical issues in developing speech systems using epoch-based analysis.

2. Review of epoch extraction methods

2.1 Epoch extraction from electroglottography

Electroglottography (EGG) is a noninvasive method of measuring the vocal fold contact dur-
ing voicing without affecting speech production (Abberton et al 1989; FFabre 1957; Fourcin &
Abberton 1971; Frokjaer-Jensen 1967; Frokjaer-Jensen & Thorvaldsen 1968). The electroglot-
tograph (EGG) measures the variation in impedance to a very small electrical current between a
pair of electrodes placed across the neck, as the area of contact of the vocal folds changes during
voicing. The demodulated impedance signal is referred to as EGG signal. During voiced speech,
the EGG signal exhibits quasi-periodicity according to the frequency of vocal-fold vibration.
Figure 1 shows a few stylized glottal cycles of the EGG signal for a voiced speech segment.
The glottal cycle of the EGG signal can be divided into four distinct phases: Closing phase,
closed phase corresponding to the region of maximum contact, opening phase and open phase.
As long as the glottis is open, the impedance measure across the larynx is maximum, and almost
flat (region-4 in figure 1). When the glottis closes, the laryngeal impedance decreases, and the
EGG signal shows a steep downward slope (region-1 in figure 1). The opening of the glottis, on
the other hand, happens much more gradually (region-3 in figure 1). It should be noted that some
authors invert the EGG signal from that shown in figure 1.
According to the theory of voice excitation (Van Den Berg 1958; Stevens 1977), the instant
of glottal closure is the point of maximum excitation to the vocal-tract system, and it is used
as the starting point of a pitch period. Although the instant of glottal closure is the most abrupt
event, it nevertheless needs a finite amount of time. The definition of the starting point of the
period, however, requires identification of a unique point in time, that is less subjected to errors.

4
Impedence

1 3

2
0 10 20 30
Time (ms)

Figure 1. EGG signal for a segment of voiced speech taken from a continuous utterance. Four distinct
phases of a glottal cycle in the EGG signal can be identified: closing phase (1), closed phase (2) with
maximum contact, opening phase (3) and open phase (4) (Murty 2009, p. 8).
 Epoch-based analysis of speech signals 655

Identifying a unique point directly from the speech waveform is not possible. But such a feature
is well manifested in the EGG signal (Scherer et al 1988). Since the EGG signal measures
directly the laryngeal impedance, it is not affected by the ambient noise. The point of inflection
during the steep fall of the EGG signal, i.e., the instant of maximum change of the laryngeal
impedance, is typically selected to represent the instant of glottal closure (Lecluse 1977). Hess &
Indefrey (1987) defined an epoch to occur at the maximum of the time-derivative of the smoothed
EGG signal during a glottal cycle. Huckvale (2000) developed an algorithm that identifies epoch
locations as the positive-going zero-crossings in the smoothed time-derivative of the EGG signal.
Figure 2 shows a segment of voiced speech, its EGG signal and the differenced EGG signal.
The locations of sharp negative peaks in the differenced EGG signal denote the instants of glottal
closure. The negative peak amplitude of the differenced EGG signal denotes the maximum flow
declination rate, which can be hypothesized to be the strength of the epoch. It can be noticed that,
in contrast to the speech signal, the EGG signal is hardly affected by the time-varying vocal-
tract system. Also changes in the shape and amplitude of the cycles in EGG are relatively small.
Hence, the epoch locations and their strengths can be accurately determined from the EGG signal
even in the dynamic regions where the vocal-tract system is not stationary.
Since every glottal cycle is represented by a single pulse, the EGG signal can be used for
accurate determination of instantaneous fundamental frequency of the voiced speech segments.
In addition, the EGG signal provides the basis for a good voiced–unvoiced discrimination, since
the differenced EGG signal is almost zero during unvoiced segments, where the glottis is always
open. Since the EGG signal is not normally available in practice, there exists strong motivation

Speech signal

0.5

0 (a)
−0.5

0.24 0.25 0.26 0.27 0.28

EGG signal
0.5

0 (b)

−0.5
0.24 0.25 0.26 0.27 0.28

Differenced EGG signal

0.05

−0.05 (c)
−0.1

0.24 0.25 0.26 0.27 0.28

Time (s)

Figure 2. Extraction of epoch locations from differenced EGG signal. (a) A segment of voiced speech
taken from a continuous utterance, (b) EGG signal, and (c) differenced EGG signal. Locations of the
negative peaks in the differenced EGG signal correspond to the instants of glottal closure (Murty 2009,
p. 9).
656 B Yegnanarayana and Suryakanth V Gangashetty

to develop techniques for extracting these source features from the speech signal. Some of the
techniques presented in the literature are reviewed below.

2.2 Epoch extraction from speech signals

Several methods have been proposed for estimating the instants of glottal closure from speech
signal without using the EGG signal. For convenience, these methods are categorized as follows:
(i) methods based on short-time energy of the speech signal, (ii) methods based on predictability
of an all-pole linear predictor, and (iii) methods based on properties of group-delay. It should be
noted that the methods placed in one category could also belong to another, depending on the
interpretation of the method.

2.2a Epoch extraction from short-time energy of speech signal: Glottal closure instants can be
detected from the energy peaks in waveforms derived directly from the speech signal (Jankowski
Jr et al 1995; Ma & Willems 1994) or from the features in its time–frequency represen-
tation (Navarro-Mesa et al 2001; Tuan & d’Alessandro 1999). In Ananthapadmanabha &
Yegnanarayana (1975), a method based on the composite signal decomposition was proposed for
epoch extraction of voiced speech. A superposition of nearly identical waveforms was referred
to as a composite signal. The epoch filter proposed in this work, computes the HE of the highpass
filtered composite signal to locate the epoch instants. It was shown that the instants of excitation
of the vocal tract could be identified precisely even in continuous speech. However, this method
is suitable for analysing only clean speech signals.
The Frobenius norm offers an estimate of the short-term energy of the speech signal. The
Frobenius norm computed using a sliding window gives an estimate of the energy value at
every speech sample. The locations of the peaks in the energy signal indicate the glottal clo-
sure instants. Frobenius norm approach based on singular value decomposition was proposed in
Ma & Willems (1994). The method was shown to work only for vowel segments. No attempt
was made to detect epochs for difficult cases such as nasals, voiced consonants and semivowels,
as in these cases, due to loading of the glottis by the vocal tract system, the strength of excitation
at the epoch is not strong as in the case of vowels.
The energy peaks can also be detected in a time–frequency representation of the speech
signal. Wavelet transform has been used to represent the speech and to detect the glottal clo-
sure instants (Tuan & d’Alessandro 1999). Lines of amplitude maxima in the time–frequency
plane were identified, and the epochs were determined using the line carrying the maximum
accumulated amplitude within each pitch period. Alternatively, a Cohen’s class time–frequency
representation of speech was constructed and used to detect the epochs (Navarro-Mesa et al
2001). The epochs were detected as peaks in a spectral density correlator derived from the
time–frequency representation.

2.2b Epoch extraction from LP: Many methods for epoch extraction rely on the discontinuities
in a linear model of speech production. An early approach used a predictability measure to detect
epochs by finding the maximum of the determinant of the autocovariance matrix of the speech
signal (Sobakin 1972; Strube 1974).
The error signal obtained in the LP analysis, referred to as the LP residual, is known to con-
tain information pertaining to epochs. A large value of the LP residual within a pitch period
is supposed to indicate the epoch location (Atal & Hanauer 1971). However, epoch identifica-
tion directly from the LP residual is not recommended (Strube 1974), because the LP residual
contains peaks of random polarity around the epochs. Figure 3b shows the LP residual derived
 Epoch-based analysis of speech signals 657

through a 10th -order LP analysis of the speech segment shown in figure 3a. The epoch loca-
tions cannot be unambiguously identified from the LP residual shown in figure 3b, because of
the occurrence of multiple peaks of either polarity. The reference epoch locations shown by the
differenced EGG signal is given in figure 3d.
A method for unambiguous identification of epochs was proposed in Ananthapadmanabha &
Yegnanarayana (1979) by computing the amplitude envelope of the analytic signal of the LP
residual, referred to as the Hilbert envelope (HE) of the LP residual. Figure 3c shows the HE of
the LP residual signal shown in figure 3b. Unlike the LP residual, the HE shows sharp unam-
biguous peaks which are in close agreement with the reference epoch locations, as shown by the
differenced EGG signal in figure 3d. Although this method works well on clean signals, perfor-
mance of the method degrades under noisy conditions due to sensitivity of the LP analysis to
noise in the signal.
Covariance analysis in the least squares approach for accurately performing the glottal inverse
filtering of the acoustic speech waveform is used in Wong et al (1979). In this work, epochs
were detected based on a measure derived from the total energy of the LP residual derived
over a sliding window. This method was further enhanced in Plumpe et al (1999), using the
observation that the formant modulations are slower in the closed phase region than in the open
phase region (Ananthapadmanabha & Fant 1982).
One of the difficulties in using the prediction error for epoch extraction is that it often contains
effects due to resonances of the vocal-tract system, as the derived inverse filter does not com-
pletely suppress the formant frequencies. As a result, the excitation peaks become less prominent
in the residual signal. In an attempt to overcome this limitation, a method based on maximum
likelihood theory for epoch determination is proposed in Cheng & O’Shaughnessy (1989). In
this method, the speech signal was processed to get the maximum-likelihood epoch detection
(MLED) signal. The strongest positive pulse in the MLED signal indicates the epoch location
within a pitch period. The MLED signal creates not only a strong and sharp pulse at epoch,

Speech signal
0.2
0 (a)
−0.2
0.58 0.585 0.59 0.595 0.6
LP residual
0.1
0 (b)
−0.1
0.58 0.585 0.59 0.595 0.6
HE of the LP residual
0.2

0.1 (c)
0
0.58 0.585 0.59 0.595 0.6
Differenced EGG signal
0.05
0
−0.05 (d)
−0.01
0.58 0.585 0.59 0.595 0.6
Time (s)

Figure 3. HE of LP residual of the speech signal for epoch extraction. (a) A segment of voiced speech
signal taken from a continuous utterance. (b) LP residual. (c) HE of the LP residual. (d) Differenced EGG
signal for observing the reference epoch locations.
658 B Yegnanarayana and Suryakanth V Gangashetty

but also a set of weaker pulses that represent suboptimal candidates for epoch within a pitch
period. Hence, a selection function was derived using the speech signal and its Hilbert trans-
form, that emphasized the contrast between the epoch and the suboptimal pulses. The limitation
of this method is the choice of the window for deriving the selection function, and also the use
of threshold for deciding the epochs.
Kalman filtering has been applied to detect the closed phase regions in voiced speech
(McKenna 2001). The boundaries of the closed phase, i.e., the instant of glottal closure and the
instant of glottal opening, are detected using the logarithm of the determinant of the error covari-
ance matrix of the Kalman filter. This measure assesses the predictability of the speech signal,
and is able to detect the glottal closure instants, but the timing accuracy is poor.

2.2c Epoch extraction using group-delay functions: A method for detecting the epochs in a
speech signal using the properties of minimum phase signals and group-delay function was pro-
posed in Smits & Yegnanarayana (1995). The method is based on the fact that the average value
of the group-delay function of a signal within an analysis frame corresponds to the location of
the significant excitation. An improved method based on the computation of the group-delay
function directly from the speech signal was proposed in Yegnanarayana & Smits (1995). The
average value of the group-delay function is called the phase slope value, and it is computed for
each sample shift to obtain the phase slope function. The instants of zero crossings of the phase
slope function correspond to epochs. Robustness of the group-delay based method against addi-
tive noise and channel distortions was studied in Murty & Yegnanarayana (1999). Four measures
of group-delay (average group-delay, zero-frequency group-delay, energy weighted group-delay
and energy weighted phase) and their use for epoch detection were investigated in Brookes et al
(2006). The effect of the length of the analysis window, the trade-off between the detection rate
and the timing error, and the computational cost of evaluating the measures were also examined
in detail. It was shown that the energy weighted measures performed better than the other two
measures.

2.3 DYPSA algorithm for epoch extraction

A dynamic programming projected phase-slope algorithm (DYPSA) for automatic estimation of
glottal closure instants in voiced speech was presented in Kounoudes et al (2002) and Naylor et al
(2007). The candidates for GCI were obtained from the zero-crossings of the phase-slope func-
tion derived from the energy weighted group-delay, and were refined by employing a dynamic
programming algorithm. It was shown that DYPSA performed better than the existing methods.
Epoch is an instant property. However, in most of the methods discussed above (except the
group-delay-based methods), the epochs are detected by employing block processing of signal
or its LP residual, which may result in ambiguity about the precise locations of the epochs.
Moreover, the excitation impulses need not be periodic, although some of these methods exploit
periodicity for accurate estimation of the epoch locations. In general, it is difficult to detect the
epochs in the case of low-voiced consonants, nasals and semivowels, breathy voices and female
speakers. A new method which overcomes some of these limitations is discussed in the next
section.

2.4 Epoch extraction using zero-frequency resonator

The discontinuities in the excitation signal caused by the sharp closure of the glottis can be
approximated by a sequence of impulses of varying amplitudes. The effect of the impulse-like
 Epoch-based analysis of speech signals 659

excitation is reflected across all frequencies, including the zero-frequency (0 Hz). The effect of
the discontinuity due to the impulse-like excitation is clearly visible in the output of narrowband
filtering (i.e., near ideal resonator at any frequency) of the speech signal. The advantage of
choosing the zero-frequency (0 Hz) resonator is that the characteristics of the time-varying vocal-
tract system will not affect the characteristics of the discontinuities in the output of the resonator.
This is because the vocal-tract system has resonances at much higher frequencies than at the
zero-frequency. An ideal zero-frequency digital resonator is an infinite impulse response filter
with a pair of poles located on the unit circle. The decay in the frequency response is determined
by the digital approximation of the all-pole filter. A cascade of two such resonators is used to
provide sharper cut-off to reduce the effect of resonances of the vocal-tract system. The following
steps are involved in processing speech signal to derive the zero-frequency filtered signal (Murty
& Yegnanarayana 2008; Yegnanarayana et al 2008).
(i) The speech signal s[n] is differenced to remove any slowly varying component introduced
by the recording device.
x[n] = s[n] − s[n − 1] (1)

(ii) The differenced speech signal x[n] is passed through a cascade of two ideal zero-frequency
(digital) resonators, i.e.,


4
y0 [n] = − ak y0 [n − k] + x[n], (2)
k=1

where a1 = −4, a2 = 6, a3 = −4 and a4 = 1. The resulting signal y0 [n] grows

approximately as a polynomial function of time.
(iii) The average pitch period is computed using the autocorrelation function of 30 ms segments
of x[n].
(iv) The trend in y0 [n] is removed by subtracting the local mean computed over the average
pitch period, at each sample. The resulting signal

1 
N
y[n] = y0 [n] − y0 [n + m] (3)
2N + 1
m=−N

is the zero-frequency-filtered (ZFF) signal. Here, 2N + 1 corresponds to the number of samples

in the window used for trend removal. The choice of the window size is not critical as long
as it is in range of one to two pitch periods. Figure 4b shows the ZFF signal of the speech
segment shown in figure 4a. It was shown that the instants of positive-to-negative zero crossings
(PNZCs) correspond to the instants of significant excitation in voiced speech, or epochs (Murty
& Yegnanarayana 2008). The locations of the PNZCs in the ZFF signal are shown by downword
arrows in figure 4c. There is close agreement between the locations of the strong positive peaks
of the negative of the differenced electroglottograph (DEGG) signal and the instants of PNZCs
derived from the ZFF signal. Information about the strength of impulse-like excitation can also
be derived from the ZFF signal, using the slope of the ZFF signal around PNZCs (Murty et al
2009). The slope is measured by computing the difference between the negative and positive
sample values on either side of the epoch, and is denoted as strength of impulse () (Murty et al
2009). Figure 4d shows the plot of , derived from the ZFF signal in figure 4b. The plot of 
shows a trend similar to the DEGG signal (figure 4c).
660 B Yegnanarayana and Suryakanth V Gangashetty
Speech signal
1

0 (a)
0 0.1 0.2 0.3 0.4 0.5
ZFR signal
400

0 (b)
−400
0 0.1 0.2 0.3 0.4 0.5
Differenced EGG signal and epoch locations
1.5

(c)
−0.2
0 0.1 0.2 0.3 0.4 0.5
Strength of excitation

30
15 (d)
0
0 0.1 0.2 0.3 0.4 0.5

Figure 4. (a) Segment of a speech signal from VOQUAL’03 (d’Alessandro & Scherer 2003) database,
(b) zero-frequency filtered signal, (c) negative of the differenced EGG signal and epoch locations marked
by arrows, and (d) strength of excitation of impulse calculated from the zero-frequency filtered signal.

It is interesting to note that even for an aperiodic sequence of impulse-like excitations, the
positive zero-crossings of the ZFF signal correspond to the locations of the epochs (Murty 2009).
There is no such relation between the excitation and the ZFF signal for random noise excitation
of the time-varying all-pole system. Further, the ZFF signal has significantly lower amplitudes
for the random noise excitation compared to the impulse sequence excitation (Murty 2009). This
is the case for unvoiced segments of speech.

2.5 Comparison of epoch extraction methods

In this section, the zero-frequency-based epoch extraction method is compared with three exist-
ing methods in terms of identification accuracy and robustness against degradation. The three
methods chosen for comparison are the HE-based method (Rao et al 2007), the group-delay
(GD)-based method (Smits & Yegnanarayana 1995) and the DYPSA algorithm (Naylor et al
2007).
The CMU-Arctic database (Kominek & Black 2004) is used to evaluate the proposed method
of epoch extraction, and to compare the results with the existing methods. In addition to the

Table 2. Performance comparison of epoch extraction methods on CMU-Arctic database.

IDR – Identification rate, MR – Miss rate, and FAR – False alarm rate, and IDA – Identification
accuracy (Murty & Yegnanarayana 2008).

Method IDR (%) MR (%) FAR (%) IDA (ms)

HE-based 89.86 1.43 8.71 0.58
GD-based 92.80 4.01 3.18 0.67
DYPSA 96.66 1.76 1.58 0.59
Zero-frequency-based 99.04 0.18 0.77 0.36
 Epoch-based analysis of speech signals 661
Table 3. Performance comparison for epoch detection methods for various SNRs and noise environments.
IDR – Identification rate (from Murty 2009, p. 67).

Environment IDR (%)

Noise SNR (dB) HE-based GD-based DYPSA Zero-frequency method
White 20 84.56 87.34 92.12 99.04
White 15 82.26 84.65 85.33 99.06
White 10 79.45 81.07 75.95 99.05
Babble 20 86.73 89.45 96.42 99.02
Babble 15 84.88 87.27 96.14 98.99
Babble 10 82.51 84.32 95.48 98.83
HF Channel 20 84.23 86.54 95.89 99.04
HF Channel 15 82.04 83.87 94.99 99.05
HF Channel 10 79.24 80.13 92.4 99.06
Vehicle 20 89.75 92.67 96.67 99.06
Vehicle 15 89.58 92.49 96.6 98.93
Vehicle 10 89.25 92.18 96.64 97.83

speech signals, the Arctic database contains simultaneous recordings of EGG signals. Reference
locations of the epochs were extracted from the voiced segments of the EGG signals by finding
peaks in the negative of the differenced EGG signal. The performance of the algorithms was
evaluated only in the voiced segments (detected from EGG signal) between the reference epoch
locations and the estimated epoch locations. The database contains a total of 792249 epochs in
the voiced regions.
Table 2 shows the performance results on Arctic database for identification rate, miss rate,
false alarm rate and identification accuracy for the three existing methods, HE-based, GD-based
and DYPSA algorithm, as well as for the zero-frequency method. From table 2, the zero-
frequency method of epoch extraction gives better identification rate as well as identification
accuracy, compared to all the three methods.
The effect of noise on the accuracy of epoch detection methods is evaluated using artifi-
cially generated noisy speech data. Several noise environments at varying SNR were simulated
to evaluate the robustness of the epoch detection methods. Table 3 shows the comparative per-
formance of epoch extraction methods for different types of degradations at varying SNRs. The
zero-frequency method consistently performs better than the existing techniques even under
degradation.

3. Epoch-based speech analysis

In this section, speech analysis methods are discussed to illustrate the significance of epochs.

3.1 Glottal activity detection

The primary mode of excitation in speech production is due to vibration of vocal folds at the
glottis, and hence may be considered as glottal activity. The strength of impulses during the
glottal activity is determined mostly by the rate of closure of the vocal folds in each glottal cycle.
662 B Yegnanarayana and Suryakanth V Gangashetty

In the absence of vocal fold vibration, the vocal-tract system may be excited by random noise, as
in the case of frication. The energy of the random noise excitation is distributed both in time and
frequency domains. Whereas the energy of an impulse is highly concentrated in the time domain,
but is distributed uniformly in the frequency domain. The ZFF signal exhibits significantly lower
amplitudes for random noise excitation compared to the impulse-like excitation, and hence can
be used to detect the regions of glottal activity (vocal fold vibration) as illustrated in figure 5.
Figure 5a shows a segment of speech signal with regions of glottal activity, marked by dotted
lines, obtained from the differenced EGG signal in figure 5b. The ZFF signal of speech, shown
in figure 5c, clearly indicates the regions of glottal activity, and the regions match well with
those obtained from the differenced EGG signal in figure 5b. It can be noticed that the unvoiced
regions around 0.6 s and 1.2 s in the speech signal (figure 5a) have very low amplitude in the
ZFF signal (figure 5c). Hence, the short-term energy of the ZFF signal computed over 20 ms
frames, shown in figure 5d, can be used for glottal activity detection (GAD).
The above GAD method was evaluated under different noisy environments (Varga &
Steeneken 1993) at varying levels of degradation. A subset of CMU-Arctic database (Kominek
& Black 2004) consisting of 100 randomly selected sentences from each of the three speakers
was used to evaluate the GAD method. The entire dataset was sample-wise labelled for glottal
activity using the simultaneously recorded EGG signals available with the database.
The performance of the proposed GAD method was evaluated using the detection error trade-
off (DET) curves (Martin et al 1997), which show the trade-off between false alarm rate (FAR)
and false rejection rate (FRR). The performance of the system is expressed in terms of equal
error rate (EER), the point at which FAR and FRR are equal. The lower the EER value, the
higher is the accuracy of the GAD method. Figure 6 shows the DET curves obtained for the GAD
algorithm under different noise environments at an SNR of 0 dB. The performance of GAD
at varying levels of degradation is listed in table 4 using the reference derived from the EGG
signals.

Speech signal
0.5

0 (a)
−0.5
0 0.2 0.4 0.6 0.8 1 1.2 1.4
Differenced EGG signal
0

−0.5 (b)
GA GA GA GA GA GA
−1
0 0.2 0.4 0.6 0.8 1 1.2 1.4
ZFR signal
0.5

0 (c)
−0.5
0 0.2 0.4 0.6 0.8 1 1.2 1.4
Energy of ZFR signal
1

0.5
(d)
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4
Time (s)

Figure 5. Glottal activity detection from the filtered signal. (a) Speech signal. (b) Differenced EGG
signal. (c) Zero-frequency filtered signal. (d) Energy computed over 20 ms segments of the filtered signal.
Regions marked as GA in (b) indicate regions of glottal activity (Murty 2009).
 Epoch-based analysis of speech signals 663

40

20
False rejection rate (%)
o

o
10

5 cle o
o
an
da
ta

1
1 2 5 10 20 40
False alarm rate (%)

Figure 6. DET curves indicating the performance of GAD method under different noise environments
(Murty 2009). Circles in figure correspond to EERs.

3.2 Estimation of instantaneous fundamental frequency

The duration of each glottal cycle is the fundamental period, and the reciprocal of this period is
referred to as the fundamental frequency. Fundamental frequency estimation is often associated
with voicing decision to eliminate the unvoiced regions from considerations. The glottal activity
detection discussed in section 3.1 is used to detect the voiced regions. Since the glottal activity
detection is based on the strength of the impulse in each glottal cycle, the end-points of the
voiced regions can be obtained accurately.
To reduce the effects due to spurious zero-crossings in the filtered signal, the HE of speech
signal is also used. The HE contains a sequence of strong impulses around the glottal clo-
sure instants, and may also contain some spurious impulses at other places due to the formant
structure of the vocal tract, and the secondary excitations in the glottal cycles. However, the
amplitudes of the impulses around the glottal closure instants dominate over those of the spurious
impulses in the computation of the ZFF signal. Hence, the ZFF signal of the HE mainly contains

Table 4. Performance of GAD in EER (%) under different noise environments

at varying levels of degradation. Reference is derived from EGG signals (Murty
2009).

Noise type 20 dB 15 dB 10 dB 5 dB 0 dB
White 3.56 3.56 3.60 3.78 5.24
Babble 3.56 3.64 4.62 7.95 15.10
Vehicle 3.56 3.58 4.09 6.28 10.83
664 B Yegnanarayana and Suryakanth V Gangashetty

the zero-crossings around the instants of glottal closure. However, the zero-crossings derived
from the ZFF signal of HE are not as accurate as those derived from the ZFF signal of speech sig-
nal. Hence, the accuracy of the zero-crossings derived from the filtered signal of speech, and the
robustness of the zero-crossings derived from the HE are used in conjunction to obtain an accu-
rate and robust estimate of the instantaneous fundamental frequency (Yegnanarayana & Murty
2009).
The instantaneous pitch frequency contour obtained from the ZFF signal of speech is used
as the primary pitch contour, and the errors in the contour are corrected using the pitch contour
derived from the HE of the speech signal. The pitch frequency contours are obtained for every 10
ms from the zero-crossings of the ZFF signals. The value of 10 ms is chosen for comparison with
the results from other methods. Let ps [m] and ph [m] be the pitch frequency contours derived,
respectively, from the speech signal and the HE of the speech signal. The following logic is used
to correct the errors in ps [m]:

ph [m], if ps [m] > 1.5 ph [m]
p[m] = (4)
ps [m], otherwise,

where m is the frame index for every 10 ms, and p[m] is the corrected pitch contour. The factor
1.5 is used mainly to reduce the pitch doubling errors in ps [m] due to spurious zero-crossings.
Any value between 1.3 and 1.8 is adequate to perform this correction.
Results of the above epoch-based method of extracting the instantaneous F0 are compared
with the results from four other methods available in the literature, in terms of accuracy in estima-
tion and robustness against degradation. The four methods are: (i) Praat’s autocorrelation method
(Boersma 1993), (ii) crosscorrelation method (Goldberg & Riek 2000), (iii) subharmonic sum-
mation (SHS) (Hermes 1988), and (iv) a fundamental frequency estimator (YIN) (de Cheveigne
& Kawahara 2002). The performance of the epoch-based method and the other four methods is
evaluated on two sets of databases: Keele database (Plante et al 1995) and the CSTR database
(Bagshaw et al 1993). Table 5 gives the performance of pitch contours derived from the five
methods. The percentage gross errors for the epoch-based method are significantly lower than
the percentage gross errors for other methods.
The effect of noise on the accuracy of the pitch estimation algorithms is examined for three
different types of noises (white Gaussian noise, babble noise and vehicle noise) collected from

Table 5. Performance of algorithms for fundamental frequency estimation on

clean data. ps [m] denotes the pitch contour derived from filtered speech signal.
ph [m] denotes the pitch contour derived from filtered HE. p[m] denotes the
pitch contour obtained by combining evidence from ps [m] and ph [m] as in
Eq. 4 (Murty 2009).

Keele Database CSTR Database

Method GE (%) M (Hz) SD (Hz) GE (%) M (Hz) SD (Hz)
AC 5.345 2.656 3.694 5.238 4.777 6.820
CC 6.891 2.201 3.371 6.818 5.108 6.730
YIN 3.219 2.165 2.906 3.073 4.922 6.584
SHS 10.774 1.868 2.398 8.938 4.108 5.864
ps [m] 2.935 3.198 4.555 3.394 5.459 6.974
ph [m] 5.647 4.562 6.381 4.157 5.699 6.886
p[m] 2.603 3.207 4.473 1.943 5.367 6.801
 Epoch-based analysis of speech signals 665

Noisex-92 database (Varga & Steeneken 1993). The results for different SNRs given in Murty
(2009) establish the robustness of the epoch-based method which can be attributed to the
impulse-like nature of the glottal closure instants in the speech signal, as all the other methods
depend mostly on the periodicity of the signal in successive glottal cycles. It should be noted
that the periodicity of the signal waveform is affected by noise.

3.3 Estimation of formant frequencies

Owing to time varying vocal-tract system during production of speech, the resonances (formants)
of the vocal tract change continuously with time. In fact, even within a glottal cycle there is
significant change in the size and shape of the vocal tract in the closed and open phase of the
glottis. The epoch locations help in locating these regions, as the region immediately after the
epoch corresponds mostly to the closed phase of the glottis, where the formants correspond to
the size and shape of the vocal tract above the glottis. These formants are also stronger and
sharper (lower bandwidth) compared to the formants in the open glottis region, which is the
region before the epoch. Although epoch locations help to separate the open and closed glottis
regions, it is a challenge to extract the formant frequencies from short (<3 ms) segments of the
closed glottis region of the speech signal.
The characteristics of the phase of Fourier transform can be exploited for extracting the for-
mant frequencies from short segments of speech (Yegnanarayana 1978). The problem of phase
wrapping can be overcome by using group delay function, which is defined as the negative
derivative of the phase of the Fourier transform of the signal (Joseph et al 2006). The group delay
function τ (ω) for a resonator with magnitude response |H (ω)| is directly proportional to the
square of the magnitude response around the resonance frequency ω0 , i.e., τ (ω0 ) ∝ |H (ω0 )|2
(Yegnanarayana 1978). Further, the group delay function for a cascade of resonators is the sum
of the group delays of the individual resonators. The high resolution and the additive proper-
ties of the group delay function are exploited to derive the formant frequencies directly from the
speech signal. The group delay function τ (ω) is computed from the speech signal x[n] using the
following relation (Guruprasad 2010):
X I (ω)Y R (ω) − X R (ω)Y I (ω)
τ (ω) = . (5)
X 2R (ω) + X 2I (ω)

Here, X (ω) = X R (ω)+ j X I (ω) is the Fourier transform of x[n], while Y (ω) = Y R (ω)+ jY I (ω)
is the Fourier transform of y[n] = − jnx[n]. The computation of the group delay func-
tion is affected by the zeros in the magnitude spectrum (|X 2 (ω)| = X 2R (ω) + X 2I (ω)). On the
other hand, if we consider only the numerator term of the group delay function, i.e., g(ω) =
X I (ω)Y R (ω) − X R (ω)Y I (ω), then g(ω) ∝ |X (ω)|4 , since τ (ω) ∝ |X (ω)|2 near the resonance
frequency (Joseph et al 2006). This indicates that the numerator of the group delay function has
higher resolution than the group delay function itself. This property of the group delay function
can be exploited to derive formant frequencies from short segments of speech signal at every
epoch. Figure 7 shows the extracted formants from the peaks of g(ω) computed at each epoch. It
should be noted that, here the formant contours are derived in a nonparametric way, as no model
(as in linear prediction) is assumed to extract the formants.
Since the regions around the epochs also correspond to high SNR regions within each glottal
cycle, it is also possible to extract the formants even from degraded speech, such as speech
collected at a distance (say 6 feet) from the speaker, even though the distant speech signal is
degraded by noise and mild reverberation. This is possible, provided the epochs are extracted
666 B Yegnanarayana and Suryakanth V Gangashetty

0 (a)

−1
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
1

0 (b)

−1
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
Frequency (Hz) Frequency (Hz)

4000

2000 (c)

0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8

4000

2000 (d)

0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8

Time (s)

Figure 7. (a) A segment of close-speaking speech signal and (b) the corresponding distant-speaking
speech signal. Contours of formants derived from (c) the close-speaking speech signal and (d) the distant
speech signal (Guruprasad & Yegnanarayana 2010).

from distant speech (Guruprasad & Yegnanarayana 2010). Figure 7 shows the formant contours
from distant (6 feet) speech signal. It is observed that the formant contours are similar to those
obtained from the close speaking speech signal.

3.4 Glottal activity in stop consonant analysis

The primary and most important mode of excitation is due to the activity at the glottis. In normal
voiced excitation (called modal voicing), there will be vibrations of the vocal folds resulting in
glottal opening, followed normally by an abrupt closure of the vocal folds, and then a closing
phase of the glottis, before the glottis is opened again for the next cycle due to build-up of
pressure from the lungs. Other aspects of glottal activity include vibration with large opening
for production of breathy voice, a complete opening for the production of unvoiced sounds, a
partial closure of the vocal folds for production of creaky voice, and finally a complete closure
of the vocal folds such as for glottal stops. Figure 8 illustrates the continuum of phonation types
as proposed by Gordon & Ladefoged (2001). In this subsection, the focus is on extraction of the
excitation characteristics due to glottal activity, in order to derive the acoustic correlates of stop
consonants using the excitation information.

Most open Most closed

Phonation type: Voiceless Breathy Modal Creaky Glottal closure

Figure 8. Phonation types (Gordon & Ladefoged 2001).

 Epoch-based analysis of speech signals 667

Stop consonants are a class of speech sounds whose characteristic feature is an interval during
which the airflow is completely blocked within the oral cavity. The air pressure built up behind
the oral closure is released more or less impulsively as the vocal tract moves towards a configu-
ration appropriate for the following vowel. Depending on the place of closure in the oral cavity,
different linguistic contrasts of the stop consonants can be produced. A stop consonant is said
to be voiced (in a phonetic sense) if there is an audible laryngeal pulsation during the closure
phase, and unvoiced if it is absent. Another phonetic feature traditionally attributed to the stop
consonants is aspiration.
Figure 9 illustrates the relative placement of the important events in the production of var-
ious stop consonants. The interval between the time of burst release to the time of onset of
vocal fold vibration is called the voice onset time (VOT) (Ambramson & Lisker 1965). The
VOT is a commonly used feature to analyse the manner of articulation in the production of stop
consonants.
The two modes of excitation in the production of stop consonants are: (i) vocal fold vibration,
and (ii) burst and aspiration. The regions of vocal-fold vibration can be extracted from the zero-
frequency filtered signal. The burst and aspiration regions correspond to band-limited noise, and
can be analysed using the LP residual. The normalized error η(n) is defined as the ratio of the
energy of the LP residual and the speech signal for a block or frame of data of size 20 ms centred
around the sample at the instant n. The normalized error is computed for every sample shift of
the frame. Figure 10 illustrates the excitation source features for the four velar stop consonants
/ka/, /kh a/, /ga/, /gh a/ in Indian languages. The plots of the filtered output clearly shows the
regions of glottal activity.
The following observations can be made to distinguish the four categories:
(i) Unvoiced unaspirated: There is sudden increase in the normalized error at the release of the
burst. The normalized error is large in the short burst region relative to the modal voicing
region.
(ii) Unvoiced aspirated: There is sudden increase in the normalized error at the release of the
burst. The large η[n] is extended over the aspirated region due to the presence of breathy

Burst−release Voicing Burst−release Voicing

Frication Frication

Aspiration

(a) Unvoiced unaspirated (b) Unvoiced aspirated

Voicing Burst−release Voicing Burst−release

Aspiration

Frication Frication

(c) Voiced unaspirated (d) Voiced aspirated

Figure 9. Schematic representation of the important events in the stop consonants (Murty 2009).
668 B Yegnanarayana and Suryakanth V Gangashetty

1 B V B V
(a) Speech signal 1
/ka/ h (d) Speech signal
/k a/
0 0

0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
1 1
(b) Filtered output (e) Filtered output
0 0

0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
1 1
(c) Normalized error (f) Normalized error
0.5 0.5

0 0
0.2 0.25 0.3 0.35 0.4 0.15 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Time (s) Time (s)

(a) /ka/ (b) /k h a/

V B 1 V B
1 (j) Speech signal
/ga/ (g) Speech signal /g ha/
0 0

0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
1 1
(h) Filtered output (k) Filtered output
0 0

0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
1 1
(i) Normalized error (l) Normalized error
0.5 0.5

0 0
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Time (s) Time (s)
h
(c) /ga/ (d) /g a/

Figure 10. The speech signal, filtered (ZFF) output, and the normalized error for four different velar stop
sound units (Murty 2009).

noise. The η[n] is low in the modal voicing region. The ZFF output is somewhat less
periodic in the aspirated region.
(iii) Voiced unaspirated: There is relatively large output in the ZFF signal due to initial voicing
compared to the relatively small amplitude in the waveform. There is increase in the η[n]
during the short burst region.
(iv) Voiced aspirated: The ZFF output is large during the initial voicing region, and then in the
aspirated and modal voicing regions. There is a dip in the ZFF output at the burst release.
However, the η[n] has an abrupt raise at the burst release, followed by large η[n] in the
aspirated region due to breathy noise.
Similar features are observed in stop consonants produced at other places of articulation. The
burst release instant (marked as B) is determined as the instant where there is a large increase in
η[n]. The starting instant of the glottal activity (marked as V) is derived from the ZFF output.
The interval between these two instants can be used as VOT to characterize the different stop
consonants.

3.5 Determination of phonetic features of glottalization, creaky voice and glottal stop
For producing normal pulmonic stop consonants (discussed in section 3.4), the articulators form
an oral closure, building up pressure behind the closure by the air coming from lungs, and the
built-up pressure is released abruptly by oral opening. The glottis is wide open for voiceless
pulmonic stops, and the vocal folds at the glottis vibrate for voiced pulmonic stops. On the other,
glottal stop and glottalized (ejective) sounds are produced by closed glottis.
 Epoch-based analysis of speech signals 669

In the production of glottalized sounds, in addition to some form of constriction (narrowing or

closing the air passage) in the vocal-tract system, the excitation air is not pulmonic (not coming
from lungs). The larynx with closed glottis moves upward like a piston to generate the excitation
air stream. This is the point where the glottalized sounds differ from the pulmonic sounds for the
place of articulation. Amharic (the official language of Ethiopia) has five ejective sounds: four
stops, /p’/, /t’/, /Ù’/, /k’/ and one dental fricative /s’/. Each of these five sounds has corresponding
voiced and voiceless pulmonic conjugates (Leslau 1995), i.e., sounds with the same place of
articulation. Table 6 gives the list of ejective sounds in Amharic along with their corresponding
voiced and voiceless pulmonic conjugates according to the place of articulation. In the case of
ejective fricative (/s’/), there is an oral constriction (producing frication noise) instead of the oral
closure as in the case of ejective stops. Since the acoustic characteristics of ejective sounds differ
from the corresponding voiced and voiceless pulmonic sound conjugates, mainly in the source of
excitation, epoch-based analysis is useful, in addition to the spectral or spectrographic analysis.
The most frequently used features for analysis of ejective sounds are: VOT, closure duration,
overall duration, BSE (burst spectral entropy), ART (amplitude rising time) of the post-target
vowel and mean cross-correlation coefficients (MXCC). To measure the VOT and closure dura-
tion of the ejective sounds and other voiceless pulmonic stops, it is necessary to determine the
instant of burst release and the onset of glottal activity. It is difficult to obtain these instants pre-
cisely from the waveform or from the spectrogram (wide-band or narrow-band) due to burst and
frication noise, and also due to background noise. The precise location of the onset of the glot-
tal activity is obtained using the zero-frequency filter output, the instant of burst release can be
located using the normalized residual computed from the LP analysis, which is relatively low
in the silence region than in the burst region. The place of articulation, however, is obtained by
using the spectral entropy measured at the output of the burst.
The spectral entropy and MXCC are computed as follows:
Calculate the spectral energy E[k] of a frame of speech using
 N −1 2
 
 − j2π 
E[k] =  x[n]e N kn  , (6)
 
n=0

where N is the length of the analysis frame, and k = 0, 1, 2, 3, ..., K − 1 are the sample points
in the frequency domain.
The spectral entropy P of a frame of speech is given by
−1
 

K
E[k] E[k]
P=−  K −1 log  K −1 , (7)
k=0 l=0 E[l] l=0 E[l]

where K is the number of samples in the frequency domain.

Table 6. Ejective sounds along with their corresponding voiced and voiceless
pulmonic conjugates in Amharic. (* Stop sounds found only in loan words.)

Manner Stops Fricative

Labial Dental Alveolar Velar Dental
Voiced /b/ /d/ /Ã/ /g/ /z/
Voiceless /p/∗ /t/ /Ù/ /k/ /s/
Ejective /p’/∗ /t’/ /Ù’/ /k’/ /s’/
670 B Yegnanarayana and Suryakanth V Gangashetty

The mean cross-correlation coefficient γ of the LP residual signal e[n] between two successive
cycles for the first three pitch periods of a vowel.
 M−1
1
3
m=0 e[m]e[m + Nk + l]
γ =   M−1 2 , (8)
3 M−1 2
k=1 m=0 e [m] m=0 e [m + N k + l]

where M = min(Nk , Nk+1 ), and Nk and Nk+1 are the number of samples in successive cycles,
l is the small relative shift in number of samples that may be necessary to get the maximum
cross-correlation coefficient, and k = 1, 2, 3.
Figures 11 and 12 show the spectrogram, waveform, ZFF signal and the normalized residual
of the LP analysis for a word with voiceless pulmonic stop and ejective, respectively. From the
figures it is obvious that the burst release can be detected from the normalized residual plot by its
prominence in amplitude, and the onset of glottal activity from the ZFF signal by the appearance
of periodicity. The time interval between the burst release and the onset of glottal activity is the
VOT (marked as V in figures 11c and 12c). The closure duration (marked as C) is measured from
the last significant excitation of the preceding vowel to the burst release.
These parameters can also be derived from the waveform and spectrogram, as marked in
figures 11a and 12a. We call these parameters as those derived from spectral methods, whereas
the parameters derived from the normalized residual and the ZFF signal are referred to as derived
from nonspectral methods. It was shown that the nonspectral methods are more accurate than
the spectral methods (Worku 2010).
It is found that ejective stops have less duration of frication than their voiceless pulmonic
conjugates. This may be because the amount of air for excitation is limited to the volume of the
oral cavity, unlike in the voiceless consonants, where the air comes from the lungs. As can be
seen in figures 11 and 12, the amplitude rising time of the vowel following the ejective sounds is
longer than for the vowel following the voiceless pulmonic stops.

(a) Spectogram

(b) Waveform

(c) Normalized residual and the filtered signal

(d) Pitch period

Figure 11. Illustration for voiceless pulmonic stops. (a) Wide-band spectrogram, (b) waveform, (c) fil-
tered (ZFF) signal and normalized residual of the LP analysis, and (d) pitch period contour for the part
(boldface) of the word /z@k@z@k@/ (Worku 2010).
 Epoch-based analysis of speech signals 671

(a) Spectogram

(b) Waveform

(c) Normalized residual and the filtered signal

(d) Pitch period

Figure 12. Illustration for ejectives. (a) Wide-band spectrogram, (b) waveform, (c) filtered (ZFF) signal
and normalized residual of the LP analysis, and (d) pitch period contour for the part (boldface) of the word
/z@k’@z@k’@/ (Worku 2010).

All ejective sounds are found to be followed by a creaky (irregular periodicity) vowel, which
is the result of dynamics of the vocal folds to change its state from complete tight closure to a
relaxed vibration or normal phonation (Stevens & Hajek 2004; Vassiére 1997). Creaky phonation
is characterized by having high or irregular pitch period (Dilley et al 1996; Gordon & Ladefoged
2001). As can be seen from figures 11c and 12c, the pitch contour of the vowel following /k’/
begins with more irregular pattern than the vowel following /k/. Irregularity of a signal can also
be measured by the cross-correlation coefficients of the waveforms in successive pitch periods.
The patterns of glottal stop are different from the supra-laryngeal consonants, in the sense that
the features of the vowel preceding the glottal stop spread across it into the following vowel,
whereas the spreading does not occur for oral stop consonant (Borroff 2007). If the two flanking
vowels are different, then the formant transitions are caused by the place of the oral consonant
(C), whereas the transition is caused by the following vowel in the V1 PV2 for the glottal stop.
Therefore, the glottal stop information cannot be inferred from the formant contours.
The voice source is aperiodic, and somewhat irregular, in successive periods during the glottal
stop region, compared with voice source in the vowel region. The aperiodicity is measured from
the successive positive zero crossing intervals in the ZFF output of the LP residual. The normal-
ized cross-correlation coefficient of successive segments is used as a measure of dissimilarity
between successive segments.
The dissimilarity is less in the vowel regions compared to the dissimilarity in the glottal stop
region (figure 13d). The successive glottal pulse intervals in the vowel regions is more steady
compared to the intervals plot in glottal stop region (figure 13e). It should be noted that the
formant contours are steady in the glottal stop region also (figure 13f), which shows that the
vocal-tract system parameters do not show the presence of the glottal stop.
Creaky voice is a result of laryngealization (where the vocal folds are held stiff and vibration
is partially inhibited). Laryngealization can cause the arytenoid cartilages and ligament portion
alternating with high and low amplitudes, which is perceived as creaky sound. The irregularity
in the pitch period contour and the low values of the normalized cross-correlation coefficient
clearly bring out the nature of the creaky voice as shown in figure 14.
672 B Yegnanarayana and Suryakanth V Gangashetty
1
Signal
0
(a)
−1

1
(b)
0
LP residual and zero crossings of filtered signal
−1

1
Filtered signal
(c)
0

−1

1
Correlation coefficient
0.5
(d)

15
Pitch period contour
Frequency (kHz) T0 (ms)

10
(e)
5

4
Formant contours (f)
2

0
0.1 0.15 0.2 0.25 0.3 0.35
Time (s)

Figure 13. Analysis of glottal stop in VPV context: (a) speech signal, (b) LP residual and zero crossings
of filtered (ZFF) signal, (c) filtered (ZFF) signal derived from LP residual, (d) normalized correlation
coefficient, (e) contour of pitch period (T0 ) and (f) formant contours (Worku 2010).

3.6 Characterization of loudness of speech

Several measures of loudness of speech have been proposed in the literature, based on physiolog-
ical characteristics of speech production and acoustic characteristics of speech signal (Monsen
& Engebretson 1977; Gauffin & Sundberg 1989). Acoustic correlates of loudness include the
features of the glottal wave (such as closed quotient, open quotient and closing quotient), and
spectral features derived from the speech signal (such as spectral tilt, measure of spectral energy
in the high-frequency region relative to that in the low-frequency region, harmonic richness, and
sharpness of spectral peaks at formant frequencies). It is observed from the EGG signals of soft
and loud speech utterances, that the abruptness of closing phase in the glottal cycle is higher in
the loud speech, compared to that in the soft speech. This indicates that the impulse-like nature
of the glottal excitation plays an important role in the perception of loudness of speech. The
objective is to derive a feature in the acoustic speech signal, which represents the impulse-like
nature of glottal excitation.
Figure 15 shows segments of voiced speech, chosen from soft, normal and loud utterances
of a speaker. The impulse-like excitation, as observed from the LP residuals of the speech seg-
ments, is more spread out in time for the soft utterances (figure 15d), compared to the normal
(figure 15e) and the loud (figure 15f) utterances. The impulse-like nature of the glottal excitation
can be observed clearly from the HE of the LP residual, shown in figures 15g, h, and i for soft,
normal, and loud utterances, respectively.
 Epoch-based analysis of speech signals 673
1

0.5

0
(a)
−0.5
Signal
−1
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55

0.5

0
(b)
−0.5
Filtered signal and its zero crossings
−1
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55

1
Correlation coefficient
0.8
(c)
0.6

0.4

0.2
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55

15
Pitch period contour

(d)
T0 (ms)

10

0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55
Time (s)

Figure 14. Analysis of creaky utterance ‘had’. (a) speech signal, (b) filtered (ZFF) signal derived from
LP residual, and its zero crossing points, (c) cross-correlation coefficients of successive pitch period of the
LP residual and (d) pitch period (T0 ) contour (Worku 2010).

A measure of the strength of excitation can be derived from a short segment (about 3 ms)
around the instant of impulse-like excitation. To represent the sharpness in the HE of the LP
σ
residual, a feature called strength of excitation is defined as η = μ , where μ denotes the mean of
the samples of the HE of the LP residual in a segment around the instant of impulse-like excita-
tion, and σ denotes the standard deviation of the samples. The ratio between standard deviation
and mean is also defined in statistics, where it is known as the coefficient of variation. Figure 16
shows an example of the distribution of η for two female and two male speakers, for soft and
loud utterances. From figures 16a, c and d, it is observed that the distribution of η does discrim-
inate between the soft and loud utterances of the speakers. The degree of discrimination is not
significant in figure 16b. The degree of discrimination, or separation between the distributions
of η for soft and loud utterances, is speaker dependent.

3.7 Analysis of Lombard effect speech

The speech produced by a person depends on several factors, which include the environment
and the auditory self-feedback of the speech of his/her own voice. Adverse environment not only
674 B Yegnanarayana and Suryakanth V Gangashetty

1 1 1
(a) (b) (c)

0 0 0

−1 −1 −1
10 20 30 10 20 30 10 20 30
1 1 1
(d) (e) (f)

0 0 0

−1 −1 −1
10 20 30 10 20 30 10 20 30
1 1 1
(g) (h) (i)
0.5 0.5 0.5

0 0 0

10 20 30 10 20 30 10 20 30
Time (ms) Time (ms) Time (ms)

Figure 15. Illustration of the nature of excitation in soft, normal and loud utterances. Speech segments in
(a), (b), and (c) belong to soft, normal, and loud utterances, respectively. The segments correspond to the
vowel ‘/a/’ in the word ‘party’ in the sentence, ‘She has left for a great party today’. Figures (d), (e), and (f)
show the LP residual for the signals in (a), (b), and (c), respectively, while the figures (g), (h), and (i) show
the HE of the corresponding LP residuals (Guruprasad & Yegnanarayana 2009).
Normalized frequency
Normalized frequency

0.4 (a) 0.4 (b)

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 0.5 1 1.5 0 0.5 1 1.5
Strength of excitation Strength of excitation
Normalized frequency
Normalized frequency

0.4 (c) 0.4 (d)

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 0.5 1 1.5 0 0.5 1 1.5
Strength of excitation Strength of excitation

Figure 16. Distribution of the strength (η) of excitation for four speakers. In each case, the broken and
the solid lines correspond to soft and loud utterances, respectively. Figures (a) and (b) correspond to two
female speakers, while (c) and (d) correspond to two male speakers (Guruprasad & Yegnanarayana 2009).
 Epoch-based analysis of speech signals 675

corrupts the speech signals by additive noise, but they also affect the self-feedback of the speech
to the person. Lack of self-feedback also affects the articulatory movement in the speech pro-
duction process, resulting in speech which the listener perceives as not normal. The speaker tries
to adjust the articulatory and acoustic parameters to produce speech as intelligence as possible
to the listeners. This psychological effect on speaker producing speech in the presence of noise
is termed as Lombard effect (Lombard 1911). The Lombard effect not only affects the intelli-
gibility in speech communication, but it also affects the performance of automatic speech and
speaker recognition systems.
It is interesting to analyse the Lombard effect speech in terms of features of excitation source
in speech production, as the time varying excitation changes significantly under the influence of
external feedback. The excitation features considered are the fundamental frequency F0 (pitch),
the strength of the impulse () at the epoch and a measure of loudness. An increase in F0 is
observed for the Lombard effect speech compared to the normal speech. The strength of impulse
decreases for Lombard effect speech compared to normal speech.
The effect of increased loudness due to Lombard effect cannot be observed in the distribution
of the values of the loudness measure (η) proposed in section 3.6. The perceived loudness of
the Lombard effect speech not only depends on the sharpness of the peaks in the HE around
the epochs, but also on the fundamental frequency (F0 ). A measure of perceived loudness (β) is
proposed to reflect this (Bapineedu 2010). The new measure β is given by.
β = η × F0 . (9)
Figure 17 shows the β contours of normal speech and for Lombard effect speech, which show
the increased loudness for Lombard effect speech in some regions.

3.8 Analysis of laugh signals

Laughter is a nonverbal vocalization that occurs often in continuous speech. The vocalized
expression of laughter varies across gender, individuals and context. Despite its variability,
laughter is perceived naturally by humans. Although laughter has very distinct perceivable

160

120
(a)
β

80

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8

160

120
(b)
β

80

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8

Time(s)

Figure 17. β contours for (a) normal speech, and (b) Lombard effect speech, for an utterance of the
sentence, ‘Regular attendance is seldom required’ (Bapineedu 2010).
676 B Yegnanarayana and Suryakanth V Gangashetty

pattern, its production is not guided by any rules of grammar as in the case of speech. It is typ-
ically produced by series of sudden bursts (outflows) of air, keeping the vocal tract in a neutral
position. The main difference between speech and laughter is that normal speech does not dis-
rupt breath, whereas laughter may (Wallace 2007). In normal speech, major role is played by the
articulators of the vocal-tract system, whereas in the case of laughter, major role is played by the
lungs and the vocal folds.
Laughter can be broadly divided into two types based on its glottal activity: (i) voiced laughter
and (ii) unvoiced laughter. In a voiced laughter, the air bursts flow mostly through the mouth,
and in some cases it may even pass through the nose. There will be more air pressure build up
in lungs, as a result of which the vocal folds may vibrate in a different way. Further, since there
is more air flow, there will be turbulence generated within the vocal folds, and hence the signal
may be breathy when compared to normal speech. In unvoiced laughter there is no voicing, the
durations of the calls will be less, and there will be more damping when compared to voiced
laughter.
Since the vocal tract is in the neutral position while laughing, we hypothesize that most of
the laughter characteristics occur due to variations in the excitation. The laughter signals are
analysed at two levels: (i) call level and (ii) bout level. Based on this kind of analysis, the features
can be categorized into two groups: call level features and bout level features. The call level
features are used to capture the call level patterns (variations within a call) of laughter, whereas
bout level features are used for capturing the high-level repetitive structure (patterns between
calls) of laughter.
The source and system characteristics of laugh signals at call level are analysed using features
such as pitch period (T0 ), strength of impulse () at epochs, amount of breathiness, call dura-
tions and some parameters derived from them. It is observed that pitch frequency for laughter is
more than that for normal speech (Bachorowski et al 2001). The general pattern that is observed
in the pitch period within a call is that it starts with some value, decreases slightly to some min-
imum, and then increases rapidly to a high value as shown in figure 18b. The main issue here is
extracting the pitch period accurately.
Since there is large amount of air pressure build up in the case of laughter, (as large amounts
of air is exhaled), the closing phase of the vocal folds is very fast. This will result in an increase
in the strength of impulse () at epoch. Figure 18c shows this general trend of  in the calls
within a bout. The pitch period contour of laughter has a unique pattern of rising rapidly at the
end of a call. So, the slope of the pitch period contour is used to capture this pattern.
As in the case of the pitch period, the strength of impulse at epochs also changes rapidly. It
rises rapidly to some maximum value and again falls at the same rate. Hence, the slope of the
normalized strengths are used to capture this pattern. Using the features in the excitation at call
and bout levels, it is possible to spot laughter segments in continuous speech as shown in Kumar
(2010).

3.9 Analysis of speech produced at different speaking rates

When humans modify speaking rate, they do not perform a simple expansion or compression
of the speech signal. To maintain the intelligibility and naturalness of the speech, they mod-
ify some of the characteristics of the speech production mechanism in a complex way. This
causes the acoustic features extracted from the speech signal to change in a complex way. These
changes affect the performance of speech systems such as speech recognition and speaker recog-
nition. Most of the studies on the effect of speaking rate on acoustic features focus on features
at segmental and suprasegmental levels. Epoch-based analysis is useful to study the effect of
 Epoch-based analysis of speech signals 677

0 (a)
−1
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75
T (ms)

5 (b)
0

0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75

0.5 (c)
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75
Time (s)

Figure 18. (a) Waveform of a voiced laughter. (b) Pitch period contour derived from epoch locations. (c)
Strength of impulses at epochs derived from the zero-frequency filtered signal (Kumar 2010).

speaking rate on features at the subsegmental level. The features are instantaneous F0 , strength
of impulse at epoch, perceived loudness, and their distributions at different speaking rates. For
this study, the speech material consists of 10 English sentences uttered by 25 male speakers at
three different speaking rates, namely, fast, normal and slow.
Figure 19 shows the distributions of instantaneous F0 for four male speakers, chosen at ran-
dom from the set of 25 speakers. The distributions of the instantaneous F0 for four speakers
are examined to illustrate differences among individuals, indicating the speaker-specific nature
of these variations. It is observed that the distribution of instantaneous F0 does discriminate
between fast, normal, and slow utterances of the speakers, although the amount of discrimination
is speaker-dependent. For speakers in figures 19a and b, there is good discrimination between
distributions of the instantaneous F0 for fast, normal and slow utterances. Discrimination can
be observed from the mean of the distributions, and from their spreads. For the speaker in fig-
ure 19c there is very little difference between distributions of fast and normal utterances, but
some discrimination can be seen between the distributions of slow and normal utterances. The
distributions shown in figure 19d are very close to each other. Some speaker-specific character-
istics can be inferred from the distributions shown in figure 19. For a speaker with a naturally
fast speaking rate, the distinction between instantaneous F0 of his/her fast and normal speech
will be less. Similarly, for a speaker with a naturally slow speaking rate, the instantaneous F0 of
slow and normal speech are similar. Some speakers are able to produce speech at three different
speaking rates while maintaining intelligibility and naturalness. In most cases, speech uttered in
at least one of the non-normal (fast or slow) speaking rates showed significant difference from
the speech uttered in normal and the other nonnormal (slow or fast) speaking rate.
Figure 20 shows the distributions of  for the four male speakers (same as used in figure 19).
From figure 20, it is observed that  does vary with speaking rate. The degree of variation is
speaker-dependent. The general trend observed across all the speakers is that ¯ (where ¯ denotes
the mean value of ) of fast speech is less than that of normal and slow speech. For some speak-
ers, ¯ of normal speech is less than that of slow speech (figures 20b and d), whereas for some
678 B Yegnanarayana and Suryakanth V Gangashetty

0.8 0.8

Normalized frequency

Normalized frequency
0.6 0.6

0.4 0.4

0.2 0.2

0 0
60 120 180 240 60 120 180 240
(a) F (Hz) (b) F (Hz)
0 0

0.8 0.8
Normalized frequency

Normalized frequency
0.6 0.6

0.4 0.4

0.2 0.2

0 0
60 120 180 240 60 120 180 240
(c) F0(Hz) (d) F0(Hz)

Figure 19. Distributions of instantaneous F0 for four male speakers. In each case, the solid (‘—’), the
dashed (‘- - -’), and the dotted (‘· · · ’) lines correspond to normal, fast and slow utterances, respectively
(Reddy 2010).

others ¯ of slow speech is less than that of normal speech (figures 20a and c). It should be noted
that this is a speaker-specific property. The spread of distribution of the values of  for fast speech
is less than that of normal and slow speech. This implies that the variation of  is less for fast
speech compared to slow and normal speech. For speakers in figures 20a, b and c, there is good
discrimination between the distributions of  of fast, normal, and slow speech. For the speaker
in figure 20d, the discrimination is very less.
It is observed that the loudness measure η varies very little with speaking rate, even though the
perceptual loudness scores show that fast speech is perceived to be louder than normal speech
(Reddy 2010).

3.10 Processing of speech collected at a distance

Speech signal collected at a distance (>50 cm) from a speaker is significantly different from
that collected close (5–10 cm) to the mouth of the speaker, due to reduction in the amplitude
of the direct component with distance, and due to effects of reverberation and noise. Hence,
extraction of fundamental frequency of voicing from distant speech signals is a challenging task.
The robustness of the impulse-like excitations in voiced speech can be exploited to extract the
fundamental frequency from distant speech signals. The key idea is that the impulse-like excita-
tions in the direct component are relatively stronger than those in the reverberant components.
 Epoch-based analysis of speech signals 679

0.5 0.5

Normalized frequency

Normalized frequency
0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
(a) ε (b) ε

0.5 0.5
Normalized frequency

Normalized frequency
0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
(c) ε (d) ε

Figure 20. Distributions of strength of excitation () for four male speakers. In each case, the solid (‘—’),
the dashed (‘- - -’), and the dotted (‘· · · ’) lines correspond to normal, fast and slow utterances, respectively
(Reddy 2010).

The speech signal is filtered through a cascade of resonators located at zero-frequency (Murty
& Yegnanarayana 2008). The ZFF signal preserves the information specific to the fundamental
frequency of speech. Figure 21a shows a voiced segment of distant speech signal. The corre-
sponding ZFF signal is shown in figure 21b. The information of the pitch periodicity is seen
more clearly in the ZFF signal (figure 21b) than in the distant speech signal (figure 21a), despite
the presence of spurious zero crossings in the ZFF signal. The ZFF signal is free from the effects
of resonances of the vocal tract. Even though the ZFF signal (figure 21b) consists of some spu-
rious zero crossings, it is observed that the fundamental frequency is the strongest component
in the short-time spectrum of the ZFF signal. The location F̃0 of peak the in the magnitude of
the short-time spectrum of the ZFF signal provides an estimate of the fundamental frequency.
An all-pole filter is constructed, such that the location of the pole is chosen corresponding to
F̃0 . The ZFF signal is passed through the all-pole filter, resulting in a signal which is nearly free
of the influence of spurious zero crossings in voiced regions (figure 21c). The reciprocal of the
time interval between successive positive-to-negative zero crossings in the refined ZFF signal is
hypothesized as the instantaneous fundamental frequency. Thus, this process helps in deriving
the instantaneous fundamental frequency accurately (Guruprasad & Yegnanarayana 2010).
Performance of the ZFF method is evaluated on SPEECON database, which consists of speech
signals collected in three different environments, namely, car interior, office and living rooms
(Iskra et al 2002). The speech signals are collected using a head-mounted close-talk microphone,
and simultaneously, using microphones placed at distances of 30 cm, 1 m and 2–3 m from the
speaker. These four cases are denoted by C0 , C1 , C2 and C3 , respectively. Gross error is used to
680 B Yegnanarayana and Suryakanth V Gangashetty

0.5

0 (a)

−0.5
1.35 1.36 1.37 1.38 1.39 1.4 1.41 1.42 1.43 1.44
1

0 (b)

−1
1.35 1.36 1.37 1.38 1.39 1.4 1.41 1.42 1.43 1.44
0.01

0 (c)

−0.01
1.35 1.36 1.37 1.38 1.39 1.4 1.41 1.42 1.43 1.44
Time (s)

Figure 21. (a) Speech signal collected at a distance. (b) The filtered (ZFF) signal. (c) The refined fil-
tered (ZFF) signal (Guruprasad 2010).

evaluate the accuracy of extraction of the fundamental frequency. It is defined as the percentage
of voiced frames for which the extracted value of the fundamental frequency deviates from the
reference value by more than 20%. In addition, the mean and standard deviation of the absolute
value of the difference between the extracted and the reference values of fundamental frequency
are also used for evaluation. Performance of the ZFF method is compared against those of
some state-of-the-art methods, which are based on time-domain and frequency-domain process-
ing of speech. These include autocorrelation method (Boersma 1993), crosscorrelation method
(Goldberg & Riek 2000), robust algorithm for pitch tracking (Talkin 1995), subharmonic
summation method (Hermes 1988), subharmonic-to-harmonic ratio method (Sun 2002), and fun-
damental frequency estimator (de Cheveigne & Kawahara 2002). Table 7 lists the performance
of the ZFF method and those of the existing methods, for the task of extraction of the funda-
mental frequency. The performance is listed for speech signals collected in office environment.
The ZFF method is denoted by ZFFM, which is based only on the interval between successive
impulse-like excitations. Table 7 shows that the ZFFM performs better than the existing algo-
rithms. The mean error of ZFFM is greater than that of the other methods, due to inclusion of
more frames in the computation of the mean error. The method is also robust for speech signals
collected from car interior and living room environments (Guruprasad & Yegnanarayana 2010).

3.11 Pitch extraction from multispeaker data

The signal collected by a microphone in a multispeaker environment is a mixture of speech
signals from several speakers. Pitch extraction from multispeaker speech signals is a challeng-
ing task, as the pitch periods from all the speakers overlap, making it difficult to observe the
individual pitch periods. Figure 22a and b show the speech signals collected by a pair of micro-
phones when two persons (one male and one female) are speaking simultaneously. It is difficult
to observe the pitch periods corresponding to the speakers from any of the two signals. Even
 Epoch-based analysis of speech signals 681
Table 7. Performance of the proposed method for estimation of fundamental frequency in close-speaking
and distant speech signals, collected in office environment. Performance of six existing methods is also
listed for comparison. For each distance, the least values of gross error, mean error and standard deviation
among the different methods are indicated in boldface (Guruprasad & Yegnanarayana 2010). The abbrevia-
tions in the table corresponds to methods: AC – autocorrelation, CC – crosscorrelation, SHS – subharmonic
summation, SHRP – subharmonic-to-harmonic ratio, RAPT – robust algorithm for pitch tracking, YIN –
fundamental frequency estimator, and ZFFM – zero-frequency filtering method.

GE (%) M (Hz) SD (Hz)

C0 C1 C2 C3 C0 C1 C2 C3 C0 C1 C2 C3
AC 6.67 8.14 10.28 32.74 3.12 3.32 3.37 6.09 4.77 5.10 5.29 7.49
CC 6.68 8.54 10.83 33.75 2.51 2.78 3.24 7.00 4.15 4.61 5.09 7.99
SHS 14.06 16.56 18.91 45.63 2.94 3.00 2.98 5.14 4.20 4.40 4.53 6.81
SHRP 8.98 10.96 20.15 61.17 3.32 3.38 3.37 5.43 4.54 4.70 4.74 6.86
RAPT 9.02 10.85 16.26 44.81 2.91 3.13 3.54 6.98 4.65 4.82 5.26 7.92
YIN 7.03 10.25 15.65 38.55 4.50 4.46 4.31 5.56 5.65 5.72 5.76 6.91
ZFFM 4.68 6.94 9.54 18.98 4.99 5.20 5.43 9.53 4.60 5.27 5.48 10.25

the autocorrelation of a frame (30 ms) of the two-speaker signal is not likely to yield two
unambiguous peaks corresponding to the pitch periods of both the speakers.
The characteristics of excitation around the epochs, and the robustness of the relative spacing
of the epochs in the speech signals collected at a pair of microphones can be exploited for
identifying the epoch locations corresponding to a given speaker. Let g1 [n] and g2 [n] be the
preprocessed HE (h 1 (n) and h 2 (n)) of the LP residuals of the speech signals collected at Mic-1
and Mic-2, respectively, as given in Eq. 10.

h i2 [n]
gi [n] = , i = 1, 2 (10)
n+M
1
2M+1 h i [m]
m=n−M

By aligning g1 [n] and g2 [n] after compensating for the estimated time-delay (τ̂1 ) correspond-
ing to Spkr-1, the epochs corresponding to that speaker will be in coherence, whereas the epochs
corresponding to Spkr-2 will be incoherent. By considering the minimum of g1 [n] and g2 [n− τ̂1 ],
only the preprocessed HE around the epochs corresponding to Spkr-1 are retained. It should be
noted that this operation of retaining minimum ensures that the preprocessed HE peaks at the
epochs of the other speakers are suppressed. The resulting signal is referred as the HE specific
to Spkr-1. In a similar manner, the signal that retains the HE around the epochs corresponding to
Spkr-2 can be derived. Let

h s j [n] = min(g1 [n], g2 [n − τ̂ j ]), j = 1, 2, (11)

where h s1 [n] and h s2 [n] are the signals in which the HE around the epochs corresponding to
Spkr-1 and Spkr-2, respectively, are retained.
Figure 22 illustrates the extraction of speaker-specific HE from two-speaker speech signals
collected using a pair of microphones. Figures 22c and d show the HE of the LP residuals of
the two-speaker speech signals shown in figure 22a and b, respectively. The HE consist of the
impulse-like excitations due to the epochs of both the speakers. It is difficult to separate the
682 B Yegnanarayana and Suryakanth V Gangashetty

0.5

0 (a)

−0.5
0 25 50 75 100 125 150
0.5

0 (b)

−0.5
0 25 50 75 100 125 150
1

0.5 (c)

0
0 25 50 75 100 125 150
1

0.5 (d)

0
0 25 50 75 100 125 150

0.5
(e)

0
0 25 50 75 100 125 150

0.5
(f)

0
0 25 50 75 100 125 150
Time (ms)

Figure 22. Illustration of extracting speaker-specific HE from two-speaker data collected using a pair of
microphones. Speech signal collected at (a) Mic-1 and (b) Mic-2. HE of LP residual of (c) Mic-1 signal,
and (d) Mic-2 signal. Speaker-specific HE of (e) Spkr-1 and (f) Spkr-2. The time-delays of arrival due to
Spkr-1 and Spkr-2 are 0.5 ms and −0.625 ms, respectively (Murty 2009).

peaks due to epochs of the individual speakers from any one of them. However, the speaker-
specific HE (figure 22e and f), obtained by computing the minimum of the delay-compensated
HE, clearly show epochs due to individual speakers.
The speaker-specific HE predominantly contain impulse-like excitations at the epoch locations
of the respective speakers. The pitch period of a given speaker can be estimated by measuring the
interval between two successive peaks in the speaker-specific HE of that speaker. This requires
detecting peaks from the speaker-specific HE, that contains large variation among the peak
 Epoch-based analysis of speech signals 683

amplitudes. In order to avoid the difficult task of peak detection, the ZFF approach is employed
to detect the impulse-like excitations in the speaker-specific HE. The positive zero-crossings of
the filtered HE closely match with the peaks in the speaker-specific HE as shown in figure 23.
Figure 23a and b show the speaker-specific HE of Spkr-1 and its ZFF signal, respectively. Even
the low amplitude peaks around 90 ms to 120 ms are correctly detected, while the spurious peaks
in between 50 ms to 60 ms are rightly ignored. Similar observations can be made from figure 23c
and d, which show speaker-specific HE of Spkr-2 and its ZFF signal. Hence, the zero-crossings
of the signal of the speaker-specific HE can be used to estimate the pitch of the individual
speakers.

4. Application of epoch-based analysis

In this section, some practical applications that exploit the impulse-like excitation of the vocal-
tract system and the precise locations of the epochs are described. Applications involving time

0.5 (a)

0
0 25 50 75 100 125 150

0 (b)

−1
0 25 50 75 100 125 150

0.5 (c)

0
0 25 50 75 100 125 150

0 (d)

−1
0 25 50 75 100 125 150
Time (ms)

Figure 23. Illustration of epoch extraction from speaker-specific HE using zero-frequency resonator.
(a) Speaker-specific HE of Spkr-1 and its (b) zero-frequency filtered (ZFF) signal. (c) Speaker-specific HE
of Spkr-2 and its (d) zero-frequency filtered (ZFF) signal. The arrow marks in (a) and (c) indicate the epoch
locations detected from zero-crossings of ZFF signals in (b) and (d), respectively (Murty 2009).
684 B Yegnanarayana and Suryakanth V Gangashetty

delay estimation and speech enhancement are considered in this section (Swamy et al 2007;
Yegnanarayana et al 2005). There are other applications such as pitch synchronous analysis for
speaker recognition and enhancement of throat microphone speech, which are not considered
here but reported in Joseph et al (2009), Murty et al (2008) and Reddy et al (2010).

4.1 Time delay estimation and determining number of speakers from mixed signals
One of the important problems in signal processing is to estimate the number of sources from
multisensor data. In multispeaker data, the problem is to determine the number of speakers, and
then localize and tract the speakers from the mixed signals received by the sensors. Solutions to
these problems are needed, especially for signals collected in a practical environment, such as a
room with background noise and reverberation.
In a multispeaker multimicrophone scenario, assuming that the speakers are stationary with
respect to the microphones, there exists a fixed time-delay of arrival of speech signals (between
every pair of microphones) for a given speaker. The time-delays corresponding to different
speakers can be estimated using the cross-correlation function of the multispeaker signals. Posi-
tions of dominant peaks in the cross-correlation function of the multispeaker signals give the
time-delays due to all the speakers at the pair of microphones. However, in general, the cross-
correlation function of the multispeaker signals does not show unambiguous prominent peaks
at the time-delays. This is mainly because of the damped sinusoidal components in the speech
signal due to resonances of the vocal tract, and also because of the effects of reverberation and
noise. These effects can be reduced by exploiting the characteristics of the excitation source
of speech. In particular, the speech signal exhibits relatively high SNR and high signal-to-
reverberation ratio (SRR) in the vicinity of time instants of significant excitations of the vocal
tract. Although some reflected components and noise may also contribute to some high SNR
regions, their relative positions will be different in the signals collected at the two microphones.
Hence, the coherence of the high SNR regions in the direct components of the signals at the two
microphones can be exploited for estimating the time-delay.
In order to highlight the high SNR regions in the speech signal, LP residual is derived from the
speech signal using the autocorrelation method of LP analysis (Makhoul 1975). The LP resid-
ual removes the second-order correlations among the samples of the signal, and produces large
amplitude fluctuations around the instants of significant excitation. The high SNR regions around
the GCIs can be highlighted by computing the HE of the LP residual (Ananthapadmanabha &
Yegnanarayana 1979).
The cross-correlation function of the HE of the LP residual signals derived from the multi-
speaker mixed signals is used to determine the number of speakers. Apart from the large ampli-
tudes around the instants of significant excitation, the HE also contains a large number of small
positive values, which may result in spurious peaks in the cross-correlation function. The regions
around the instants of significant excitation are further emphasized by dividing the square of
each sample of the HE by the moving average of the HE computed over a short window around
the sample (see Eq. 10).
For multispeaker signals collected using a pair of microphones, p = 2. The cross-correlation
function r12 [l] between the preprocessed HE g1 [n] and g2 [n] is computed as
 N −|k|−1
g1 [n]g2 [n − l]
r12 [l] =  n=z  N −|k|−1 2 ,
N −|k|−1
n=m g12 [n] n=m g2 [n]
l = 0, ±1, ±2, . . ., ±L , (12)
 Epoch-based analysis of speech signals 685

where m = l, k = 0 for l ≥ 0, and m = 0, k = l for l < 0, and N is the length of the

segments of the HE. Here, both the vectors are normalized to unit magnitude for every sample
shift before computing the cross-correlation. The cross-correlation function is computed over an
interval of 2L + 1 lags, where 2L + 1 corresponds to an interval greater than the largest expected
delay. The largest expected delay can be estimated from the approximate positions of the speak-
ers and microphones in the room. The locations of the peaks with respect to the origin (zero
lag) of the cross-correlation function correspond to the time-delays between the microphone sig-
nals for all the speakers. The number of prominent peaks should correspond to the number of
speakers. However, in practice, this is not always true because of the following reasons: (i) All
speakers may not contribute to voiced sounds in the segments used for computing the cross-
correlation function. (ii) There could be spurious peaks in the cross-correlation function, which
may not correspond to the delay due to a speaker. Hence, we rely only on the delay due to the
most prominent peak in the cross-correlation function. This delay is computed from the cross-
correlation function of successive frames of 50 ms duration. The delays are computed for each
frame shift of 5 ms. Since different regions of speech signal may provide evidence for the delays
corresponding to different speakers, the number of frames corresponding to each delay is accu-
mulated over the entire data. This helps in the determination of number of speakers, as well as
their respective delays. Thus, by collecting the number of frames corresponding to each delay
over the entire data, there will be large evidence for the delays corresponding to the individual
speakers.
Speech data was collected simultaneously using two microphones separated by about 1 m in a
laboratory environment, with an average (over the frequency range of 0.5–3.5 kHz) reverberation
time of about 0.5 seconds. The cross-correlation function of the HE of the LP residuals of the
multispeaker signals is used to estimate the time-delays. The percentage of frames for each delay
(in ms) for three, four, five and six speakers are shown in figure 24. The locations of the peaks
in the histograms correspond to the time-delays due to different speakers. Thus, the number of
peaks in the histogram indicates the number of speakers, and the heights of the peaks show the
relative prominence of each speaker in the conversation.

4.2 Speech enhancement in single channel case

Perceiving information from speech signals collected over severely degraded channels is a dif-
ficult task. To increase the comfort level of listening, one can process the speech signal to
reduce noise in the nonspeech regions. This is possible if we are able to identify the speech
regions. Noise characteristics may be estimated and subtracted from the degraded speech signal.
However, in real environments, noise characteristics vary over time. Hence, reliable estima-
tion of noise characteristics is a difficult task. Alternatively, characteristics of speech may be
exploited to process the degraded speech. One advantage of using the knowledge of speech is
that the characteristics of speech are more predictable compared to that of the noise components
(Yegnanarayana & Murthy 2000).
The speech-specific knowledge from the vocal-tract system or the excitation source or both
may be used for enhancement. In this study, the knowledge of the excitation source is used
for enhancement. The HE of LP residual containing information about the excitation source is
derived from the speech signal. One property of the HE of LP residual of the speech signal
collected over a severely degraded channel is that the correlation among the samples is high in
speech regions and low in nonspeech regions. Autocorrelation analysis may be performed on the
HE of the LP residual to estimate the amount of correlation among the samples. For illustration, a
30 ms frame of HE of LP residual computed from a strong-voiced segment of degraded speech,
686 B Yegnanarayana and Suryakanth V Gangashetty

20 20
(a) (b)
15 15

% of frames
10 10

5 5

0 0
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3
20 20
(c) (d)
15 15
% of frames

10 10

5 5

0 0
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3
Delay (ms) Delay (ms)

Figure 24. Percentage of frames for each delay in milliseconds for (a) three speakers, (b) four speakers,
(c) five speakers and (d) six speakers. The arrows indicate the peaks corresponding to different speakers
(Swamy et al 2007).

and its autocorrelation sequence are shown in figure 25a and b, respectively. The strength of
the first peak (after the central peak) in the autocorrelation sequence is an indication of the
level of correlation in the frame, which is high in this case. The HE of the LP residual of a 30
ms frame of a segment of a weak-voiced speech segment and its autocorrelation sequence are
shown in figure 25c and d, respectively. The strength of the peak is relatively low in this case.
Similarly, autocorrelation analysis performed for the HE of the LP residual of a 30 ms frame of
nonspeech regions of the signal is also shown in figure 25e and f, respectively. Therefore, the
autocorrelation analysis performed on frames of the HE of the LP residual for every sample shift
gives an indication of the level of speech at each sample in the degraded signal.
A 10th order LP analysis is performed on the degraded speech signal (see figure 26a) sampled
at 8 kHz, to obtain the LP residual. The HE of the LP residual is computed. Autocorrelation is
performed on the HE of the LP residual using frames of size 30 ms and frame shift of one sam-
ple. For each frame, the strength of the first peak of the autocorrelation sequence, normalized
with respect to the central peak, is noted. The normalized peak strength of the autocorrelation
sequence, computed for the HE of the LP residual is shown in figure 26c. High values in the
normalized peak strength indicate the speech regions. The normalized peak strength sequence is
suitably processed using a 500-point Hamming window. A weight function is derived from the
smoothed sequence using a nonlinear mapping function in such a way that the samples corre-
sponding to the speech regions are enhanced relative to the samples in the nonspeech regions.
The nonlinear mapping function may be given by
1
Pm = + α, (13)
1 + e−(Ps −θ)/τ
where Pm is value of the weight function value, Ps is the smoothed peak strength value (nor-
malized in the range 0–1). The parameters θ = 0.2, τ = 0.04 are the slope parameters, and
 Epoch-based analysis of speech signals 687
1 1

Ps=0.62
0.5 (a) 0.5 (b)

0
0
5 10 15 20 25 30 10 20 30 40 50
1 1

Ps=0.24
0.5 (c) 0.5 (d)

0
0
5 10 15 20 25 30 10 20 30 40 50
1 1

Ps=0.12
0.5 (e) 0.5 (f)

0
0
5 10 15 20 25 30 10 20 30 40 50
Time(ms) Time(ms)

Figure 25. HE of the LP residual of a 30 ms (a) strong-voiced frame and its (b) autocorrelation sequence,
(c) weak-voiced frame and its (d) autocorrelation sequence, (e) nonspeech frame and its (f) autocorrelation
sequence. Ps indicates normalized strength of the first peak (Prasanna 2004).

α = 0.05 is the offset which is the minimum value of the weight function. The weight function
derived using the mapping function is shown in figure 26d.
The LP residual of the degraded speech signal is processed using the weight function to pro-
duce the modified LP residual (figure 26e). As the samples of the LP residual signal are less
correlated compared to the samples of the speech signal, modifying the LP residual may intro-
duce less distortion in the synthesized signal. The enhanced speech signal (figure 26f) is obtained
by exciting the time-varying filter using the modified LP residual. The LP coefficients of the
filter are derived from the degraded speech signal.
From the narrowband spectrograms of the degraded and the corresponding enhanced speech
signals, we can infer that the energy of the frequency components in the speech regions are
unaltered, but the signal is attenuated significantly in the nonspeech regions (Prasanna 2004).

4.3 Multichannel speech enhancement

When speech is transmitted in an acoustical environment similar to an office room, it
will be degraded by background noise and reverberation (Boll 1979; Cheng & O’Shaughnessy
1991; Ephraim & Trees 1995; Flanagan et al 1985; Huang & Zhao 1998; Jensen & Hansen
2001; Mittal & Phamdo 2000; Miyoshi & Kaneda 1988; Nemer et al 2002; Oh & Viswanathan
1992; Satyanarayana 1999; Scalart & Benmar 1996; Silverman 1987; Subramaniam et al 1996;
Yegnanarayana et al 1997, 1999; Yegnanarayana & Murthy 2000). Multichannel case is more
effective for enhancement compared to the single channel case, but requires estimation of time-
delays (Flanagan et al 1985). A simple method for enhancement in multichannel case is addition
of the speech signals, after compensating for the delays. Coherent addition of speech sig-
nals from different microphones will provide enhancement mainly against background noise.
The improvement in enhancement is directly related to the number of microphones used. For
688 B Yegnanarayana and Suryakanth V Gangashetty

0 (a)

−1
200 400 600 800 1000 1200 1400 1600 1800 2000
1

0 (b)

−1
200 400 600 800 1000 1200 1400 1600 1800 2000
1

0.5 (c)

0
200 400 600 800 1000 1200 1400 1600 1800 2000
1

0.5 (d)

0
200 400 600 800 1000 1200 1400 1600 1800 2000
1

0 (e)

−1
200 400 600 800 1000 1200 1400 1600 1800 2000
1

0 (f)

−1
200 400 600 800 1000 1200 1400 1600 1800 2000
Time (ms)

Figure 26. (a) Degraded speech signal, and its (b) LP residual, (c) normalized peak strength values,
(d) weight function, (e) modified LP residual, and (f) enhanced speech signal (Prasanna 2004).

achieving significant enhancement, especially due to reverberation, additional processing of the

microphone signals is required. In this section, a method for enhancement using the epoch
information is described.
Speech was collected from 14 spatially distributed microphones placed in an office room of
dimension 3m × 4m × 3m with a reverberation time of about 200 ms. The delay between every
pair of microphones is computed using the excitation source information. The coherently-added
signal, obtained after compensating for their delays is given by
1
se1 (n) = [s1 (n) + s2 (n − τ12 ) + ...... + s N (n − τ1N )], (14)
N
where τ1i is the delay in samples between mic-1 and mic-i.
 Epoch-based analysis of speech signals 689

The coherent addition reinforces speech components and thus reduces the effect of the back-
ground noise. However, the reverberant component is still present in the resulting signal. The
degree of enhancement achieved at this level depends on the number of microphones used in the
coherent addition. For instance, from the enhanced speech signals and their narrowband spectro-
grams, when signals from 2, 5, 10 and 14 microphones are added, one can observe that there is
decrease in the background noise as the number of microphones are increased (Prasanna 2004). It
is interesting to note that in the case where signals from 14 microphones are added, even though
the effect of background noise is reduced, the reverberation tails are still present in the signal. It
is necessary to process the coherently-added speech signal further to achieve enhancement with
respect to reverberation.
For each of the microphone signals, the HE of the LP residual is computed. The coherent
addition of the HE reinforces the epoch information, whereas the incoherent addition will spread
the epoch information. The coherently-added HE exhibits several interesting features. The devi-
ation among the samples of the HE is high in the voiced speech regions. Typically, voiced
speech regions in continuous speech have a minimum duration of 50 ms. Hence, by consid-
ering a block of 50 ms duration and a shift of one sample, the mean and standard deviation
of the coherently-added HE samples in each block are computed. The standard deviation val-
ues are normalized with the respective mean values. In the normalized standard deviation plot,
the deviation of HE samples is high in the speech regions. Further, the normalized standard
deviation is high in the initial portions of the voiced speech regions, and it decreases towards
the end of the voiced regions. This is because the initial parts are high SNR regions. Towards
the end of the voiced regions, the relative levels of degrading components increase, and hence
they correspond to low SNR regions. Another interesting property of the coherently-added HE
is that, the samples in each pitch period around the epochs have large deviation compared to
the samples away from the epoch. The mean, standard deviation and normalized standard devi-
ation of the coherently-added HE are computed for a block size of 3 ms and a shift of one
sample.
A weight function is derived by adding the two (long and short blocks) normalized standard
deviation values as shown in figure 27. The weighted residual is used to excite the time vary-
ing all-pole filter derived from the coherently-added signal, to synthesize the enhanced speech
signal. The clean speech, its degraded version, coherently-added signal from three microphones
and the enhanced speech, along with their narrowband spectrograms indicate that the speech sig-
nal is enhanced both with respect to background noise as well as with respect to reverberation
(Prasanna 2004, p. 70–83).

4.4 Multispeaker separation

In a multispeaker environment, the objective is to separate the speech component corresponding
to each speaker, while retaining the quality and intelligibility as much as possible. The signal
collected by a microphone in a multispeaker environment is a mixture of speech signals from
several speakers. Processing speech for enhancement in such conditions is a challenging task, as
the speech of the other speakers acts as noise, against which the speech of the desired speaker
needs to be enhanced. The difficulty in achieving this enhancement is due to the similarity of
the spectral characteristics of the speech signals from different speakers. The difficulty is further
compounded by the fact that the spectral characteristics are modified by the response of the room
and also by the background noise. The extent of degradation depends on the relative position
of the microphone with respect to the speaker, and also on the background noise. The primary
690 B Yegnanarayana and Suryakanth V Gangashetty
1

0.5 (a)

0
0 500 1000 1500 2000 2500
1

0.5 (b)

0
0 500 1000 1500 2000 2500
1

0.5 (c)

0
0 500 1000 1500 2000 2500
Time(ms)

Figure 27. (a) Normalized standard deviation plot derived using block size of 50 ms and shift of 1 sam-
ple, (b) normalized standard deviation plot derived using block size of 3 ms and shift of 1 sample, and
(c) weight function obtained by adding (a) and (b) (Prasanna 2004).

causes of degradation are room reverberation, background noise and the distance of the speaker
from the microphone.
Most of the multispeaker enhancement methods in the literature involve modification of spec-
tral features representing the vocal tract system (Barros et al 2002; Cardoso 1998; Lee &
Childers 1988; Mitchell et al 1971; Morgan et al 1997; Frazier et al 1990; Parsons 1976). They
use the knowledge of pitch to separate the individual speakers in multispeaker case. Here, a
method is described for speaker separation from speech collected over multiple microphones, by
exploiting characteristics of the strength of excitation impulses at the epochs, and the robustness
of the relative spacing of the GC events in the speech signals collected at different microphones.
The HE of the LP residual is used as a representation for the sequence of impulses corresponding
to the instants of significant excitation of the vocal tract system. When these sequences are added
coherently using the knowledge of the time-delay of each speaker, the strengths of the excita-
tion of the desired speaker are enhanced relative to the strengths of excitation of other speakers.
Using the knowledge of the enhanced speaker characteristics in the coherently-added sequence
of impulses, a weight function is derived, which in turn is used to derive a modified excitation
signal. This modified excitation signal is used to synthesize speech using the vocal tract system
characteristics derived from the degraded speech signal. Enhancement in the resulting speech is
primarily due to enhancement of the excitation characteristics, which are important perceptually
(Prasanna 2004).
The steps in processing the two microphone signals for enhancement are summarized in
table 8 (Prasanna 2004).

4.5 Prosody modification using knowledge of epoch locations

Prosody modification involves changing pitch and duration of speech without affecting the nat-
uralness. This section gives a method for prosody modification using the instants of significant
 Epoch-based analysis of speech signals 691
Table 8. Steps for speech enhancement in multispeaker environment (Prasanna 2004).

Sl. No. Description

1 Collect the speech signals (sampling frequency 8 kHz) from two speakers over two spatially
separated microphones in a live room.
2 Derive the (10th -order) LP residuals from the speech signals.
3 Compute the HE of the LP residuals.
4 Estimate the time-delays for each speaker using the cross-correlation of the HE.
5 Add the HE using the estimated time-delays to produce the coherently-added HE for
each speaker.
6 Derive the weight function using the standard deviation plots from the coherently-added HE.
7 Derive the modified LP residual signal from each microphone signal.
8 Synthesize the enhanced speech for each microphone signal.
9 Coherently add the speech signals of the desired speaker derived from both the microphone
signals.

excitation of the vocal tract system during production of speech. The instants of significant exci-
tation correspond to the epochs in the case of voiced speech, and to some random excitation in
the case of unvoiced speech.
The method for prosody manipulation makes use of the properties of the excitation source
information. The residual signal in the LP analysis is used as an excitation signal (Makhoul
1975). Successive samples in the LP residual are less correlated compared to the samples in
the speech signal. The residual signal is manipulated by using resampler either for increasing
or decreasing the number of samples required for the desired prosody modification. The resid-
ual manipulation is likely to introduce less distortion in the speech signal synthesized using the
modified LP residual and LP coefficients (LPCs). The time-varying vocal-tract system character-
istics are represented by the LPCs for each analysis frame. Since the LPCs carry the information
about the short-time spectral envelope, they are not altered in the described method for prosody
modification. LP analysis is carried out over short segments (analysis frames) of speech data to
derive the LP coefficients and the LP residual for the speech signal (Makhoul 1975).
There are four main steps involved in the prosody manipulation: (i) Deriving the instants
of significant excitation (epochs) from the LP residual signal, (ii) deriving a modified (new)
epoch sequence according to the desired prosody (pitch and duration), (iii) deriving a modified
LP residual signal from the modified epoch sequence, and (iv) synthesizing speech using the
modified LP residual and the LPCs. Figure 28 shows the block diagram indicating various stages
in prosody modification (Rao 2005).
Since the LP residual is used for incorporating the desired prosody modification, there is
no significant distortion due to resampling the residual samples both in the voiced and in the
nonvoiced regions. This is because there is less correlation among samples in the LP residual
compared to the correlation among the signal samples.

5. Practical issues in developing speech system using epoch-based analysis

The epoch-based analysis methods and applications discussed in this paper depend critically
on our ability to extract epochs from speech signals. Epoch extraction using the ZFR output
is robust against degradation caused by additive noise. However, in practice, the speech signal
is degraded due to several factors such as reverberation in distance speech, channel effects in
692 B Yegnanarayana and Suryakanth V Gangashetty

Speech signal
LPCs
LP analysis
LP residual
Epoch extraction
(Group delay analysis)

Generation of
epoch interval plot

Generation of desired
epoch interval plot
(Pitch modification)
Updating the LPCs
Generation of desired epoch
according to analysis interval plot by resampling
frame shift (modified (Duration modification)
in case of duration
modification) Generation of new sequence
of epochs using the desired
epoch interval plot

Determining the nearest

original epoch corresponding
to each of the new epochs

Copying the residual using

the nearest original epochs

Resampling the residual

according to desired
epoch interval plot

LP filter

Synthesized speech
after prosody modification

Figure 28. Block diagram for prosody modification (Rao 2005).

communication, recording devices, source coding methods used in communication and effects
due to speech from multiple speakers.
Some of the effects of degradation can be overcome by deriving the epochs from the HE of
the LP residual of the speech signal (Yegnanarayana et al 2011). In most such cases, the peaks
of the HE will be stronger and impulse-like around the epochs. Hence, the epochs can be derived
by using the ZFF on the HE of the LP residual. However, if there are too many spurious impulses
in the signal caused by echoes or reverberation, then it is not possible to derive the epochs from
such signals. In such cases, the effects of spurious impulses can be reduced by processing the
data collected from number of spatially distributed microphones. By coherently adding the HE
of the LP residual, after compensation for the time-delay between a pair of microphones, it may
be possible to enhance the strength of the peaks in HE due to direct speech at the microphone
locations. However, even then it may be difficult to extract all the epochs from the degraded
signals. The effectiveness of human listening may tempt us to believe that it may be possible to
extract all the epochs. However, human listening uses selective attention in perceiving speech,
and hence all epochs may not be used in perceived pitch and other phonetic information. In fact,
the redundancy in speech at various levels makes it possible to get the message in the signal,
 Epoch-based analysis of speech signals 693

even if some important events are missing due to degradations. Hence, determining the relevant
information at all levels, including at the epochs level is a major challenge in speech signal
processing.
Processing multispeaker mixed data to extract epochs for each speaker is a highly challenging
task, as it involves separation of speech due to individual speakers, and then extracting the epochs
for each speaker’s speech. One way of dealing with this problem is by processing the mixed data
collected at several spatially distributed microphones. As of now, the success of this approach is
limited by the fact that it is difficult to extract the HE of the LP residual of individual speakers
by compensating for time delay of the direct speech of the speaker at a pair of microphones.
Here again the human listening may be adopting selective attention at various levels to get the
message of each speaker from the mixed signals.
Most communication situations employ some form of coding the speech signal using source
coding approach. In these cases, the waveform information is sought to be preserved as much as
possible in some mean squared sense. It is likely that in such cases the epoch information may
be modified significantly, and hence it may not be possible to derive the epoch information from
the received signal. It is likely that perceptually significant component of source information
may have been lost in the coding process. This may also suggest the need for exploring source
coding methods that preserve the significant component of the excitation source information.

Material for this tutorial paper is borrowed from several theses (MS & PhD) and papers written
in the Speech and Vision Laboratory at IIT Madras, Chennai and IIIT Hyderabad over the past
several years. The authors would like to thank their colleagues and students who have contributed
to the work presented in this paper over the recent decade or so.

References

Abberton E R M, Howard D M, Fourcin A J 1989 Laryngographic assessment of normal voice: A tutorial.

Clinical Linguistics and Phonetics 3(3): 263–296
Ambramson A S, Lisker L 1965 Voice onset time in stop consonants: acoustic analysis and synthesis. Proc.
5th Int. Congr. Phonetic Sciences, Liege, A51
Ananthapadmanabha T V, Fant G 1982 Calculations of true glottal volume-velocity and its components.
Speech Commun. 1: 167–184
Ananthapadmanabha T V, Yegnanarayana B 1975 Epoch extraction of voiced speech. IEEE Trans. Acoust.
Speech Signal Process. 23(6): 562–570
Ananthapadmanabha T V, Yegnanarayana B 1979 Epoch extraction from linear prediction residual for
identification of closed glottis interval. IEEE Trans. Acoust. Speech Signal Process. 27(4): 309–319
Atal B S, Hanauer S L 1971 Speech analysis and synthesis by linear prediction of the speech wave. J.
Acoust. Soc. Am. 50(2): 637–655
Bachorowski J, Smoski M, Owren M 2001 The acoustic features of human laughter. J. Acoust. Soc. Am.
111: 1582–1597
Bagshaw P C, Hiller S M, Jack M A 1993 Enhanced pitch tracking and the processing of F0 contours
for computer and intonation teaching. Proc. European Conf. on Speech Commun. (Eurospeech), Berlin,
Germany, 1003–1006. URL http://www.cstr.ed.ac.uk/research/projects/fda/
Bapineedu G 2010 Analysis of Lombard effect speech and its application in speaker verification for
imposter detection. MS thesis, Language Technologies Research Centre, International Institute of
Information Technology, Hyderabad, India
Barros A K, Rutkowski T, Itakura F, Ohnishi N 2002 Estimation of speech embedded in a reverberant
and noisy environment by independent component analysis and wavlets. IEEE Trans. Neural Netw. 13:
888–893
694 B Yegnanarayana and Suryakanth V Gangashetty

Boersma P 1993 Accurate short-term analysis of fundamental frequency and the hormincs-to-noise ratio of
a sampled sound. Proc. Inst. Phonetic Sci. 17: 97–110
Boll S F 1979 Suppression of acoustic noise in speech using spectral subtraction. IEEE Trans. Acoust.
Speech Signal Process. ASSP-27: 113–120
Borroff M L 2007 A landmark underspecification account of the patterning of glottal stop. PhD thesis,
Stony Brook University, New York
Brookes M, Naylor P A, Gudnason J 2006 A quantitative assessment of group delay methods for identifying
glottal closures in voiced speech. IEEE Trans. Audio Speech Lang. Process. 14(2): 456–466
Cardoso J-F 1998 Blind signal separation: statistical principles. Proc. IEEE 86: 2009–2025
Cheng Y M, O’Shaughnessy D 1989 Automatic and reliable estimation of glottal closure instant and period.
IEEE Trans. Acoust. Speech Signal Process. 27: 1805–1815
Cheng Y M, O’Shaughnessy D 1991 Speech enhancement based conceptually on auditory evidence. IEEE
Trans. Signal Process. 39: 1943–1954
CMU-ARCTIC speech synthesis databases. URL http://festvox.org/cmu_arctic/index.html
d’Alessandro C, Scherer K R 2003 Voice quality: Functions, analysis and synthesis (VOQUAL’03). ISCA
Tutorial and Research Workshop, Geneva, Switzerland, http://archives.limsi.fr/VOQUAL/voicematerial.
html (last viewed 04/08/2009)
de Cheveigne A, Kawahara H 2002 YIN, a fundamental frequency estimator for speech and music.
J. Acoust. Soc. Am. 111(4): 1917–1930
Dilley L, Shattuck-Hufnagel S, Ostendorf M 1996 Glottalization of word-initial vowels as a function of
prosodic structure. J. Phonetics 24: 423–444
Ephraim Y, Van Trees H L 1995 A signal subspace approach for speech enhancement. IEEE Trans. Speech
Audio Process. 3(4): 251–266
FFabre P 1957 Un procede electrique percutane d’inscrition de l’accolement glottique au cours de la
phonation: glottographie de haute frequence. premiers resultats. Bull. Acad. Natl. Med. 141: 66
Flanagan J L, Jonston J D, Zahn R, Elko G W 1985 Computer-steered microphone arrays for sound
transduction in large rooms. J. Acoust. Soc. Am. 78(5): 1508–1518
Fourcin A J, Abberton E 1971 First applications of a new laryngograph. Med. Biol. Illus. 21: 172–182
Frazier R H et al 1990 Enhancement of speech by adaptive filtering. Proc. IEEE Int. Conf. Acoust. Speech
Signal Process. New York, NY, USA
Frokjaer-Jensen B 1967 A photo-electric glottograph. Annual Report of the Institute of Phonetics of
University of Copenhagen 2: 5–19
Frokjaer-Jensen B, Thorvaldsen P 1968 Construction of a fabre glottograph. ARIPUC 3: 1
Gauffin J, Sundberg J 1989 Spectral correlates of glottal voice source waveform characteristics. J. Speech.
Hear. Res. 32: 556–565
Goldberg R, Riek L 2000 A practical handbook of speech coders. (Boca Raton, FL: CRC Press)
Gordon M, Ladefoged P 2001 Phonation types: a cross-linguistic overview. J. Phonetics 29(4): 383–406
Guruprasad S 2010 Significance of processing regions of high signal-to-noise ratio in speech sig-
nals. PhD thesis, Department of Computer Science and Engineering, Indian Institute of Technology
Madras
Guruprasad S, Yegnanarayana B 2009 Perceived loudness of speech based on the characteristics of
excitation source. J. Acoust. Soc. Am. 126(4): 2061–2071
Guruprasad S, Yegnanarayana B 2011 Performance of an event-based instantaneous fundamental frequency
estimator for distant speech signals. IEEE Trans. Audio Speech Lang. Process. 19(7): 1853–1864
Hamon C, Moulines E, Charpentier F 1989 A diphone synthesis system based on time domain prosodic
modifications of speech. Proc. IEEE Int. Conf. Acoust. Speech Signal Process. Glasgow, 238–241
Hermes D J 1988 Measurement of pitch by subharmonic summation. J. Acoust. Soc. Am. 83(1): 257–264
Hess W, Indefrey H 1987 Accurate time-domain pitch determination of speech signals by means of a
laryngograph. Speech Commun. 6: 55–68
Huang J, Zhao Y 1998 An energy-constrained signal subspace method for speech enhancement and recog-
nition in colored noise. Proc. IEEE Int. Conf. Acoust. Speech Signal Process. Seattle, WA, USA,
377–380
 Epoch-based analysis of speech signals 695

Huckvale M 2000 Speech filing system: Tools for speech research. URL http://www.phon.ucl.ac.uk/
resource/sfs/
Iskra D, Grosskopf B, Marasek K, Van Den Heuvel H, Diehl F, Kiessling A 2002 SPEECON – speech
databases for consumer devices: Database specification and validation. Proc. Third Int. Conf. Lang.
Resources Eval. (LREC), Las Palmas, Canary Islands - Spain, 329–333
Jankowski C R Jr, Quatieri T F, Reynolds D A 1995 Measuring fine structure in speech: Application to
speaker identification. Proc. IEEE Int. Conf. Acoust. Speech Signal Process. Detroit, MI, USA, 325–328
Jensen J, Hansen J H L 2001 Speech enhancement using a constrained iterative sinusoidal model. IEEE
Trans. Speech Audio Process. 9(7): 731–740
Joseph A M, Guruprasad S, Yegnanarayana B 2006 Extracting formants from short segments using group
delay functions. Proc. Int. Conf. Spoken Language Processing, Pittsburgh, USA, 1009–1012
Joseph A M, Yegnanarayana B, Gupta S, Kesheorey M R 2009 Speaker dependent mapping for low bit rate
coding of throat microphone speech. Proc. Interspeech 2009, Brighton, UK, 1087–1090
Kominek J, Black A 2004 The CMU Arctic speech databases. Proc. 5th ISCA Speech Synthesis Workshop,
Pittsburgh, USA, 223–224
Kounoudes A, Naylor P A, Brookes M 2002 The DYPSA algorithm for estimation of glottal closure instants
in voiced speech. Proc. IEEE Int. Conf. Acoust. Speech Signal Process. vol. 11, Orlando, FL, 349–352
Kumar S K 2010 Analysis of laugh signals for automatic detection and synthesis. MS thesis, Language
Technologies Research Centre, International Institute of Information Technology, Hyderabad, India
Lecluse F L E 1977 Elecroglottography. Dissertation, Univ. of Rotterdam
Lee C K, Childers D G 1988 Cochannel speech separation. J. Acoust. Soc. Am. 83(1): 274–280
Leslau W 1995 Reference grammar of Amharic. (Wiesbaden: Otto Harrassowitz)
Lombard E 1911 Le signe de l’elevation de la voix, annals maladiers oreille. Larynx. Nez. Pharynx. 37:
101–119
Ma Y K C, Willems L F 1994 A Frobenius norm approach to glottal closure detection from the speech
signal. IEEE Trans. Speech Audio Process. 2: 258–265
Makhoul J 1975 Linear prediction: A tutorial review. Proc. IEEE, 63(4): 561–580
Martin A, Doddington G, Kamm T, Ordowski M, Przybocki M 1977 The DET curve in assessment of
detection task performance. Proc. European Conf. Speech Process. Technol. Greece, 1895–1898
McKenna J G 2001 Automatic glottal closed-phase location and analysis by Kalman filtering. Proc. 4th
ISCA Tutorial and Research Workshop on Speech Synthesis, Perthshire, Scotland
Meyer G F 1995 Keele pitch database, School of Psychology, University of Liverpool, UK. URL http:
//www.liv.ac.uk/Psychology/hmp/projects/pitch.html
Mitchell O M M, Ross C A, Yates G H 1971 Signal processing for a cocktail party effect. J. Acoust. Soc.
Am. 50: 656–660
Mittal U, Phamdo N 2000 Signal/noise klt based approach for enhancing speech degraded by colored noise.
IEEE Trans. Speech Audio Process. 8: 159–167
Miyoshi M, Kaneda Y 1988 Inverse filtering of room acoustics. IEEE Trans. Acoust. Speech Signal Process.
ASSP-36: 145–152
Monsen R B, Engebretson A M 1977 Study of variations in the male and female glottal wave. J. Acoust.
Soc. Am. 62(4): 981–993
Morgan D P, George E B, Lee L T, Kay S M 1997 Cochannel speech separation by harmonic enhancement
and supression. IEEE Trans. Speech Audio Process. 5: 407–424
Murty K S R 2009 Significance of Excitation Source Information for Speech Analysis. PhD thesis,
Department of Computer Science and Engineering, Indian Institute of Technology Madras
Murty K S R, Yegnanarayana B 2006 Combining evidence from residual phase and MFCC features for
speaker recognition. IEEE Signal Process. Lett. 13(1): 52–56
Murty K S R, Yegnanarayana B 2008 Epoch extraction from speech signals. IEEE Trans. Audio Speech
Lang. Process. 16(8): 1602–1613
Murty K S R, Khurana S, Itankar Y U, Kesheorey M R, Yegnanarayana B 2008 Efficient representation of
throat microphone speech. Proc. Interspeech 2008, Brisbane, Australia, 2610–2613
Murty K S R, Yegnanarayana B, Joseph M A 2009 Characterization of glottal activity from speech signals.
IEEE Signal Process. Lett. 16(6): 469–472
696 B Yegnanarayana and Suryakanth V Gangashetty

Murty P S, Yegnanarayana B 1999 Robustness of group-delay-based method for extraction of significant

excitation from speech signals. IEEE Trans. Speech Audio Process. 7(6): 609–619
Navarro-Mesa J L, Lleida-Solano E, Moreno-Bilbao A 2001 A new method for epoch detection based on
the Cohen’s class of time frequency representations. IEEE Signal Process. Lett. 8(8): 225–227
Naylor P A, Kounoudes A, Gudnason J, Brookes M 2007 Estimation of glottal closure instants in voiced
speech using the DYPSA algorithm. IEEE Trans. Audio Speech Lang. Process. 15(1): 34–43
Nemer E, Goubran R, Mahmoud S 2002 Speech enhancement using fourth-order cumulants and optimum
filters in the subband domain. Speech Commun. 36: 219–246
Neocleous A, Naylor P A 1998 Voice source parameters for speaker verification. Proc. Eur. Signal Process.
Conf. 697–700
Noisex-92. URL http://www.speech.cs.cmu.edu/comp.speech/Section1/Data/noisex.html
Oh S, Viswanathan V 1992 Hands-free voice communication in an automobile with a microphone array.
Proc. IEEE Int. Conf. Acoust. Speech Signal Process. San Francisco, California, 281–284
Parsons T W 1976 Separation of speech from interfering speech by means of harmonic selection. J. Acoust.
Soc. Am. 60: 911–918
Plante F, Meyer G F, Ainsworth W A 1995 A pitch extraction reference database. Proc. European Conf.
Speech Commun. (Eurospeech) Madrid, Spain, 827–840
Plumpe M D, Quatieri T F, Reynolds D A 1999 Modeling of the glottal flow derivative waveform with
application to speaker identification. IEEE Trans. Acoust. Speech Signal Process. 7: 569–586
Prasanna S R M 2004 Event Based Analysis of Speech. PhD thesis, Department of Computer Science and
Engineering, Indian Institute of Technology Madras
Rao K S 2005 Acquisition and incorporation of prosody knowledge for speech systems in Indian languages.
PhD thesis, Department of Computer Science and Engineering, Indian Institute of Technology Madras
Rao K S, Yegnanarayana B 2006 Prosody modification using instants of significant excitation. IEEE Trans.
Audio, Speech Lang. Process. 14(3): 972–980
Rao K S, Prasanna S R M, Yegnanarayana B 2007 Determination of instants of significant excitation in
speech using Hilbert envelope and group-delay function. IEEE Signal Process. Lett. 14(10): 762–765
Reddy S H M 2010 Analysis of Speech at Different Speaking Rates using Excitation Source Information.
MS thesis, Language Technologies Research Centre, International Institute of Information Technology,
Hyderabad, India
Reddy S H M, Prahallad K, Gangashetty S V, Yegnanarayana B 2010 Significance of pitch synchronous
analysis for speaker recognition using AANN models. Proc. Interspeech 2010, Makuhari, Chiba, Japan,
669–672
Swamy R K, Murty K S R, Yegnanarayana B 2007 Determining number of speakers from multispeaker
speech signals using excitation source information. IEEE Signal Process. Lett. 14(7): 481–484
Satyanarayana P 1999 Short segment analysis of speech for enhancement. PhD thesis, Indian Institute of
Technology Madras, Department of Computer Science and Engg., Chennai, India
Scalart P, Benmar A 1996 A system for speech enhancement in the context of hands-free radiotelephony
with combined noise reduction and acoustic echo cancellation. Speech Commun. 20: 203–214
Scherer R C, Druker D G, Titze I R 1988 Vocal physiology: Voice production mechanisms and functions.
(New York: Raven Press Ltd.)
Silverman H F 1987 Some analysis of microphone arrays for speech data acquisition. IEEE Trans. Acoust.
Speech Signal Process. ASSP-35(12): 1699–1712
Smits R, Yegnanarayana B 1995 Determination of instants of significant excitation in speech using group
delay function. IEEE Trans. Speech Audio Process. 3(5): 325–333
Sobakin A N 1972 Digital computer determination of formant parameters of the vocal tract from a speech
signal. Soviet Phys.-Acoust. 18: 84–90
Stevens K N 1977 Physics of laryngeal behavior and larynx models. Phonetica 34: 264–279
Stevens M, Hajek J 2004 A preliminary investigation of some acoustic characteristics of ejectives in
Waima’a: VOT and closure duration. In S Cassidy, F Cox, R Mannell, S Palethorpe (eds) Proc. Tenth
Australian Int. Conf. Speech Science and Technology, Macquarie University, Sydney, ASSTA, 277–282
Strube H W 1974 Determination of the instant of glottal closures from the speech wave. J. Acoust. Soc. Am.
56: 1625–1629
 Epoch-based analysis of speech signals 697

Subramaniam S, Petropulu A P, Wendt C 1996 Cepstrum-based deconvolution for speech dereverberation.

IEEE Trans. Speech Audio Process. 4: 392–396
Sun X 2002 Pitch determination and voice quality analysis using subharmonic-to-harmonic ratio. Proc.
IEEE Int. Conf. Acoust. Speech Signal Processing, Orlando, FL, USA, 333–336
Talkin D 1995 A robust algorithm for pitch tracking (RAPT). Speech coding and synthesis, (Amsterdam:
Elsevier Science)
Tuan V N, d’Alessandro C 1999 Robust glottal closure detection using the wavelet transform. Proc.
European Conf. Speech Processing, Technology, Budapest, 2805–2808
Van Den Berg J 1958 Myoelastic-aerodynamic theory of voice production. J. Speech Hearing 1: 227–244
Varga A, Steeneken H J M 1993 Assessment for automatic speech recognition: II. NOISEX-92:
A database and an experiment to study the effect of additive noise on speech recognition systems. Speech
Comm. 12(3): 247–251. [Online] Available: http://www.speech.cs.smu.edu/comp.speech/Section1/Data/
noisex.html
Vassiére J 1997 Phonological use of the larynx. In ISCA LARYNX-1997, Marseille, France, 115–126
Veeneman D, BeMent S 1985 Automatic glottal inverse filtering from speech and electroglottographic
signals. IEEE Trans. Signal Process. 33: 369–377
Wallace C 2007 The phonetics of laughter – A linguistic approach. Interdisciplinary Workshop on the
Phonetics of Laughter, Saarbrucken, August 4–5 Saarbruchen
Wong D Y, Markel J D, Gray A H 1979 Least squares glottal inverse filtering from the acoustic speech
waveform. IEEE Trans. Acoust. Speech Signal Process. 27: 350–355
Worku H S 2010 Acoustic characterization of glottal stop and glottalized sounds in Amharic using non-
spectral methods of speech analysis. PhD thesis, Language Technologies Research Centre, International
Institute of Information Technology, Hyderabad, India
Yegnanarayana B 1978 Formant extraction from linear prediction phase spectra. J. Acoust. Soc. Am. 63(5):
1638–1640
Yegnanarayana B, Murty P S 2000 Enhancement of reverberant speech using LP residual signal. IEEE
Trans. Speech Audio Process. 8(3): 267–281
Yegnanarayana B, Smits R L H M 1995 A robust method for determining instants of major excitations in
voiced speech. Proc. IEEE Int. Conf. Acoust. Speech Signal Process., Detroit, USA, 776–779
Yegnanarayana B, Veldhuis R N J 1998 Extraction of vocal-tract system characteristics from speech signals.
IEEE Trans. Speech and Audio Process. 6(4): 313–327
Yegnanarayana B, Avendaño C, Hermansky H, Murty P S 1997 Processing linear prediction residual for
speech enhancement. Proc. European Conf. Speech Process. Technol. Rhodes, Greece, 1399–1402
Yegnanarayana B, Avendaño C, Hermansky H, Murthy P S 1999 Speech enhancement using linear
prediction residual. Speech Commun. 28(1): 25–42
Yegnanarayana B, Prasanna S R M, Duraiswami R, Zotkin D 2005 Processing of reverberent speech for
time-delay estimation. IEEE Trans. Speech Audio Process. 13(6): 1110–1118
Yegnanarayana B, Murty K S R, Rajendran S 2008 Analysis of stop consonants in indian languages
using excitation source information in speech signal. Proc. Workshop Speech Anal. Process. Knowledge
Discovery, June 4–6 Aalborg, Denmark
Yegnanarayana B, Murty K S R 2009 Event-based instantaneous fundamental frequency estimation from
speech signals. IEEE Trans. Audio Speech Lang. Process. 17(4): 614–624
Yegnanarayana B, Prasanna S R M, Guruprasad S 2011 Study of robustness of zero frequency resonator
method for extraction of fundamental frequency. Proc. IEEE Int. Conf. Acoust. Speech Signal Process.
Prague, Czech Republic