Voice Morphing Seminar Report
Voice Morphing Seminar Report
CHAPTER 1
INTRODUCTION
training data. Although the requirement for parallel training data is often acceptable, there
are applications which require voice transformation for nonparallel training data.
Examples can be found in the entertainment and media industries where recordings of
unknown speakers need to be transformed to sound like well-known personalities.
Further uses are envisaged in applications where the provision of parallel data is
impossible such as when the source and target speaker speak different languages.
Although interpolated linear transforms are effective in transforming speaker identity, the
direct transformation of successive source speech frames to yield the required target
speech will result in a number artifacts. The reasons for this are as follows. First, the
reduced dimensionality of the spectral vector used to represent the spectral envelope and
the averaging effect of the linear transformation result in formant broadening and a loss
of spectral detail. Second, unnatural phase dispersion in the target speech can lead to
audible artifacts and this effect is aggravated when pitch and duration are modified.
Third, unvoiced sounds have very high variance and are typically not transformed.
However, in that case, residual voicing from the source is carried over to the target
speech resulting in a disconcerting background whispering effect .To achieve high quality
of voice conversion, include a spectral refinement approach to compensate the spectral
distortion, a phase prediction method for natural phase coupling and an unvoiced sounds
transformation scheme. Each of these techniques is assessed individually and the overall
performance of the complete solution evaluated using listening tests. Overall it is found
that the enhancements significantly improve.
CHAPTER 2
Speech morphing can be achieved by transforming the signal’s representation from the
acoustic waveform obtained by sampling of the analog signal, with which many people
are familiar with, to another representation. To prepare the signal for the transformation,
it is split into a number of 'frames' - sections of the waveform. The transformation is then
applied to each frame of the signal. This provides another way of viewing the signal
information. The new representation (said to be in the frequency domain) describes the
average energy present at each frequency band. Further analysis enables two pieces of
information to be obtained: pitch information and the overall envelope of the sound. A
key element in the morphing is the manipulation of the pitch information. If two signals
with different pitches were simply crossfaded it is highly likely that two separate sounds
will be heard. This occurs because the signal will have two distinct pitches causing the
auditory system to perceive two different objects. A successful morph must exhibit a
smoothly changing pitch throughout. The pitch information of each sound is compared to
provide the best match between the two signals' pitches. To do this match, the signals are
stretched and compressed so that important sections of each signal match in time. The
interpolation of the two sounds can then be performed which creates the intermediate
sounds in the morph. The final stage is then to convert the frames back into a normal
waveform.
However, after the morphing has been performed, the legacy of the earlier analysis
becomes apparent. The conversion of the sound to a representation in which the pitch and
spectral envelope can be separated loses some information. Therefore, this information
has to be re-estimated for the morphed sound. This process obtains an acoustic
waveform, which can then be stored or listened to.
CHAPTER 3
The algorithm to be used is shown in the simplified block diagram given below. The
algorithm contains a number of fundamental signal processing methods including
sampling, the discrete Fourier transform and its inverse, cepstral analysis. However
the main processes can be categorized as follows.
I. Preprocessing or representation conversion: This involves processes like signal
acquisition in discrete form and windowing.
II. Cepstral analysis or Pitch and Envelope analysis: This process will extract the pitch
and formant information in the speech signal.
III. Morphing which includes Warping and interpolation.
IV. Signal re-estimation.
The filter amplifies energy at and near formant frequencies, while attenuating
energy around anti resonant frequencies between the formants. The common method used
to extract pitch and formant frequencies is the spectral analysis. This method views
speech as the output of a liner, time-varying system (vocal tract) excited by either
quasiperiodic pulses or random noise. Since the speech signal is the result of convolving
excitation and vocal tract sample response, separating or “deconvolving” the two
components can be used. In general, deconvolution of the two signals is impossible, but it
works for speech, because the two signals have quite different spectral characteristics.
The deconvolution process transforms a product of two signals into a sum of two signals.
If the resulting summed signals are sufficiently different spectrally, they may be
separated by linear filtering.Now we present a comprehensive analysis of each of the
processes involved in morphing with the aid of block diagrams wherever necessary.
3.2 Preprocessing
This section shall introduce the major concepts associated with processing a speech
signal and transforming it to the new required representation to affect the morph. This
process takes place for each of the signals involved with the morph.
The input speech signals are taken using MIC and CODEC. The analog speech
signal is converted into the discrete form by the inbuilt CODEC TLC320AD535 present
onboard and stored in the processor memory. This completes the signal acquisition phase.
3.2.2 Windowing
A DFT (Discrete Fourier Transformation) can only deal with a finite amount of
information. Therefore, a long signal must be split up into a number of segments. These
are called frames. Generally, speech signals are constantly changing and so the aim is to
make the frame short enough to make the segment almost stationary and yet long enough
to resolve consecutive pitch harmonics. Therefore, the length of such frames tends to be
in the region of 25 to 75 milli seconds. There are a number of possible windows. A
selection is:
3.3 Morphing
3.3.1 Matching and Warping: Background theory
Figure 3.4: The match path between two signals with different located features
The match path shows the amount of movement (or warping) required in order aligning
corresponding features in time. Such a match path is obtained by Dynamic Time
Warping (DTW).
In this use of DTW, a path between two pitch contours is required. Therefore,
each feature vector will be a single value. In other uses of DTW, however, such feature
vectors could be large arrays of values. Since the feature vectors could possibly have
multiple elements, a means of calculating the local distance is required. The distance
measure between two feature vectors is calculated using the Euclidean distance metric.
Therefore the local distance between feature vector x of signal 1 and feature vector y of
signal 2 is given by 3.3. As the pitch contours are single value feature vectors, this
simplifies to 3.4.
The global distance is the overall difference between the two signals. Audio is a
time- dependent process. For example, two audio sequences may have different durations
and two sequences of the sound with the same duration are likely to differ in the middle
due to differences in sound production rate. Therefore, to produce a global distance
measure, time alignment must be performed - the matching of similar features and the
stretching and compressing, in time, of others .Instead of considering every possible
match path which would be very inefficient, a number of constraints are imposed upon
the matching process.
The cells at (i,j) and (i,0) have different possible originator cells. The path to (i, 0) can
only originate from (i-1, 0). However, the path to (i,j) can originate from the three
standard locations as shown in the figure 3.9 above.
Fig 3.7: A sample back trace array with each cell containing a number, which represents
the location of the predecessor cell in the lowest global path distance to that cell.
For the example in Figure above, the 2D array would be Figure 3.8: The sample back
trace array with the calculated path overlaid At this stage, we now have the match path
between the pitches of the two signals and each signal in the appropriate form for
manipulation. The next stage is to then produce the final morphed signal.
CHAPTER 4
MORPHING STAGE
Now we shall give a detailed account of how the morphing process is carried out.
The overall aim in this section is to make the smooth transition from signal 1 to signal 2.
This is partially accomplished by the 2D array of the match path provided by the DTW.
At this stage, it was decided exactly what form the morph would take. The
implementation chosen was to perform the morph in the duration of the longest signal. In
other words, the final morphed speech signal would have the duration of the longest
signal. In order to accomplish this, the 2D array is interpolated to provide the desired
duration. However, one problem still remains: the interpolated pitch of each morph slice.
If no interpolation were to occur then this would be equivalent to the warped cross-fade
which would still be likely to result in a sound with two pitches. Therefore, a pitch in-
between those of the first and second signals must be created. The precise properties of
this manufactured pitch peak are governed by how far through the morph the process is.
At the beginning of the morph, the pitch peak will take on more characteristics of the
signal 1 pitch peak - peak value and peak location - than the signal 2 peak. Towards the
end of the morph, the peak will bear more resemblance to that of the signal 2 peaks. The
variable l is used to control the balance between signal 1 and signal 2. At the beginning of
the morph, l has the value 0 and upon completion, l has the value 1. Consider the example
in Figure 4.6. This diagram shows a sample cepstral slice with the pitch peak area
highlighted. Figure
4.7 shows another sample cepstral slice, again with the same information highlighted. To
illustrate the morph process, these two cepstral slices shall be used.
There are three stages:
1. Combination of the envelope information;
2. Combination of the pitch information residual - the pitch
information excluding the pitch peak;
3. Combination of the pitch peak information.
We can say that that the best morphs are obtained when the envelope information
is merely cross-faded, as opposed to employing any pre-warping of features, and so this
approach is adopted here. In order to cross-fade any information in the cepstral domain,
care has to be taken. Due to the properties of logarithms employed in the cepstral analysis
stage, multiplication is transformed into addition. Therefore, if a cross-faded between the
two envelopes were attempted, multiplication would in fact take place. Consequently,
each envelope must be transformed back into the frequency domain (involving an
inverse logarithm) before the cross-fade is performed. Once the envelopes have been
successfully cross-faded according to the weighting determined by l, the morphed
envelope is once again transformed back into the cepstral domain. This new cepstral slice
forms the basis of the completed morph slice.
The pitch information residual is the pitch information section of the cepstral slice with
the pitch peak also removed by, liftering. To produce the morphed residual, it is
combined in a similar way to that of the envelope information: no further matching is
performed. It is simply transformed back into the frequency domain and cross-faded with
respect to l. Once the cross-fade has been performed, it is again transformed into the
cepstral domain. The information is now combined with the new morph cepstral slice
(currently containing envelope information). The only remaining part to be morphed is
the pitch peak area.
liftered in such a way as to align the peaks with respect to the liftered area (see Figure
4.8).
II. The two liftered cepstral slices are then transformed back into the frequency domain
where they can be cross-faded with respect to
CHAPTER 5
4.1 The whole morphing process is summarized using the detailed block diagram
shown below (figure 6.1).
In real time voice morphing what we want is to be able to morph, in real-time user
singing a melody with the voice of another singer. It results in an “impersonating” system
with which the user can morph his/her voice attributes, such as pitch, timbre, vibrato and
articulation, with the ones from a prerecorded target singer. The user is able to control the
degree of morphing, thus being able to choose the level of “impersonation” that he/she
wants to accomplish. In our particular implementation we are using as the target voice a
recording of the complete song to be morphed. A more useful system would use a
database of excerpts of the target voice, thus choosing the appropriate target segment at
each particular time in the morphing process. In order to incorporate to the user’s voice
the corresponding characteristics of the “target” voice, the system has to first recognize
what the user is singing(phonemes and notes), finding the same sounds in the target voice
(i.e. synchronizing the sounds), then interpolate the selected voice attributes, and finally
generate the output morphed voice. All this has to be accomplished in real-time.
Figure shows the general block diagram of the voice impersonator system. The
underlying analysis/synthesis technique is SMS to which many changes have been done
to better adapt it to the singing voice and to the real-time constrains of the application.
Also a recognition and alignment module was added for synchronizing the user’s voice
with the target voice before the morphing is done. Before we can morph a particular song
we have to supply information about the song to be morphed and the song recording itself
(Target Information and Song Information). The system requires the phonetic
transcription of the lyrics, the melody as MIDI data, and the actual recording to be used
as the target audio data. Thus, a good impersonator of the singer that originally sang the
song has to be recorded. This recording has to be analyzed with SMS, segmented into
“morphing units”, and each unit labeled with the appropriate note and phonetic
information of the song. This preparation stage is done semi-automatically, using a non-
real time application developed for this task. The first module of the running system
includes the real time analysis and the recognition/ alignment steps. Each analysis frame,
with the appropriate parameterization, is associated with the phoneme of a specific
moment of the song and thus with a target frame. The recognition/alignment algorithm is
based on traditional speech recognition technology, that is, Hidden Markov Models
(HMM) that were adapted to the singing voice. Once a user frame is matched with a
target frame, we morph them interpolating data from both frames and we synthesize the
output sound. Only voiced phonemes are morphed and the user has control over which
and by how much each parameter is interpolated. The frames belonging to unvoiced
phonemes are left untouched thus always having the user’s consonants.
SMS procedures to adapt them to the requirements of the impersonator system. A major
improvement to SMS has been the real-time implementation of the whole
analysis/synthesis process, with a processing latency of less than 30 milliseconds and
tuned to the particular case of the singing voice. This has required many optimizations in
the analysis part, especially in the fundamental frequency detection algorithm. These
improvements were mainly done in the pitch candidate's search process, in the peak
selection process, in the fundamental frequency tracking process, and in the
implementation of a voiced-unvoiced gate. Another important set of improvements to
SMS relate to the incorporation of a higher-level analysis step that extracts the
parameters that are most meaningful to be morphed. Attributes that are important to be
able to interpolate between the user’s voice and the target’s voice in a karaoke application
include spectral shape, fundamental frequency, amplitude and residual signal. Others,
such as pitch micro variations, vibrato, spectral tilt, or harmonicity, are also relevant for
various steps in the morphing process or to perform other sound transformation that are
done in parallel to the morphing. For example, transforming some of these attributes we
can achieve voice effects such as Tom Waits hoarseness.
This part of the system is responsible for recognizing the phoneme that is being
uttered by the user and also its musical context so that a similar segment can be chosen
from the target information. There is a huge amount of research in the field of speech
recognition. The recognition systems work reasonably well when tested in the well-
controlled environment of the laboratory. However, phoneme recognition rates decay
miserably when the conditions are adverse. In our case, we need a speaker independent
system capable of working in a bar with a lot of noise, loud music being played and not
very-high quality microphones. Moreover the system deals with singing voice, which has
never been worked on and for which there are no available databases. It has to work also
with very low delay, we cannot wait for a phoneme to be finished before we recognize it
and we have to assign a phoneme to each frame.
This would be a rather impossible/impractical problem if it was not for the fact that we
know the words beforehand, the lyrics of the song. This reduces a big portion of the
search problem: all the possible paths are restricted to just one string of phonemes, with
several possible pronunciations. Then the problem reduces to locating the phoneme in the
lyrics and placing the start and end points. We have incorporated a speech recognizer
based on phoneme-base discrete HMM's that handles musical information and that is able
to work with very low delay. The details of the recognition system can be found in
another paper of our group. The recognizer is also used in the preparation of the target
audio data, to fragment the recording into morphable units (phonemes) and to label them
with the phonetic transcription and the musical context. This is done out of real-time for a
better performance
CHAPTER 6
FUTURE SCOPE
There are a number of areas in which further work should be carried out in order
to improve the technique described here and extend the field of speech morphing in
general. The time required to generate a morph is dominated by the signal re-estimation
process. Even a small number (for example, 2) of iterations takes a significant amount of
time even to re-estimate signals of approximately one second duration. Although in
speech morphing, an inevitable loss of quality due to manipulation occurs and so less
iteration are required, an improved re-estimation algorithm is required. A number of the
processes, such as the matching and signal re-estimation are very unrefined and
inefficient methods but do produce satisfactory morphs. Concentration on the issues
described above for further work and extensions to the speech morphing principle ought
to produce systems which create extremely convincing and satisfying speech morphs.
Further extension to this work to provide the above functionality would create a powerful
and flexible morphing tool. Such a tool would allow the user to specify at which points a
morph was to start and finish the properties of the morph and also the matching function.
With the increased user interaction in the process, a Graphical User Interface could be
designed and integrated to make the package more 'user-friendly'. Such an improvement
would immediate visual feedback (which is lacking in the current implementation) and
possibly step by step guidance.
Finally, this work has used spectrograms as the pitch and voicing and spectral
envelope representations. Although effective, further work ought to concentrate on new
representations which enable further separation of information. For example, a new
representation might allow the separation of the pitch and voicing. The Speech morphing
concept can be extended to include audio sounds in general. This area offers many
possible applications including sound synthesis. For example, there are two major
methods for synthesizing musical notes. One is to digitally model the sound's physical
source and provide a number of parameters in order to produce a synthetic note of the
desired pitch. Another is to take two notes which bound the desired note and use the
principles used in speech morphing to manufacture a note which contains the shared
characteristics of the bounding notes but whose other properties have been altered to
form a new note. The use of pitch manipulation within the algorithm also has an
interesting potential use. In the interests of security, it is sometimes necessary for people
to disguise the identity of their voice. An interesting way of doing this is to alter the pitch
of the sound in real-time using sophisticated methods.
Morphing
Depending on the phoneme the user is singing, a unit from the target is selected.
Each frame from the user is morphed with a different frame from the target, advancing
sequentially in time. Then the user has the choice to interpolate the different parameters
extracted at the analysis stage, such as amplitude, fundamental frequency, spectral shape,
residual signal, etc. In general the amplitude will not be interpolated, thus always using
the amplitude from the user and the unvoiced phonemes will also not be morphed, thus
always using the consonants from the user. This will give the user the feeling of being in
control. In most cases the durations of the user and target phonemes to be morphed will
be different. If a given user’s phoneme is shorter than the one from the target the system
will simply skip the remaining part of the target phoneme and go directly to the
articulation portion. In the case when the user sings a longer phoneme than the one
present in the target data the system enters in the loop mode. Each voiced phoneme of the
target has a loop point. frame, marked in the preprocessing, non-real time stage. The
system uses this frame to loop-synthesis in case the user sings beyond that point in the
phoneme. Once we reach this frame in the target, the rest of the frames of the user will be
interpolated with that same frame until the user ends the phoneme. This process is shown
in Figure
The frame used as a loop frame requires a good spectral shape and, if possible, a
pitch very close to the note that corresponds to that phoneme. Since we keep a constant
spectral shape, we have to do something to make the synthesis sound natural. The way
we do it is by using some “natural” templates obtained from the analysis of a longer
phoneme that are then used to generate more target frames to morph with out of the loop
frame. One feature that adds naturalness is pitch variations of a steady state note sung by
the same target. These delta pitches are kept in a look up table whose first access is
random and then we just read consecutive values. We keep two tables, one with
variations of steady pitch and another one with vibrato to generate target frames. Once all
the chosen parameters have been interpolated in a given frame they are added back to the
basic SMS frame of the user. The synthesis is done with the standard synthesis
procedures of SMS
CHAPTER 7
CONCLUSION
The approach we have adopted separates the sounds into two forms: spectral envelope
information and pitch and voicing information. These can then be independently
modified. The morph is generated by splitting each sound into two forms: a pitch
representation and an envelope representation. The pitch peaks are then obtained from the
pitch spectrograms to create a pitch contour for each sound. Dynamic Time Warping of
these contours aligns the sounds with respect to their pitches. At each corresponding
frame, the pitch, voicing and envelope information are separately morphed to produce a
final morphed frame. These frames are then converted back into a time domain waveform
using the signal re-estimation algorithm. In this seminar, only one type of morphing has
been discussed - that in which the final morph has the same duration as the longest signal.
Also we discuss the case of speech morphing in this seminar. But the work can be
extended to include audio sounds as well. The longest signal is compressed and the
morph has the same duration as the shortest signal (the reverse of the approach described
here). If one signal is significantly longer than the other, two possibilities arise. However,
according to the eventual use of the morph, a number of other types could be produced.
CHAPTER 8
REFERENCES