Noise suppression in speech

Group 844

30th May 2005

Aalborg University
Institute of Electronic Systems
Fredrik Bajers Vej 7 DK-9220 Aalborg East Telephone +45 96 35 87 00

Title: Noise suppression in speech

Theme: Adaptive DSP algorithms and Advanced Signal Analysis
Project period: 1st February 2005 - 30. May 2005

Abstract
Project group:
This project develops an algorithm to suppress the
844
noise from a noise contaminated speech signal for
Group members: mobile phone users. The one microphone scenario is
Raju Muchanthula chosen with an assumption that noise is additive and
that the speech signal and noise are uncorrelated.
The spectral subtraction technique based on mini-
Supervisors: mum statistics, is a noise power spectrum estimator
Kjeld Hermansen and subtracts noise spectrum from noisy speech sig-
nal spectrum to get the estimate of noise suppressed
Ole Wolf speech spectrum. The spectral subtraction, the spec-
tral manipulation and residual manipulation meth-
ods are implemented. The proposed algorithm is
capable of tracking different noise levels, results in
Publications: 5
improvement of signal to noise ratio and perceptual
Total Pages: 50
quality and the intelligibility is maintained.
Supplement: CD-ROM
The algorithm is analyzed for real time and tested
Finished: 30/05-2005
according to objective evaluation criteria. The algo-
rithm is chosen for mapping on an architecture of
TMS320C6713 DSP processor.
 Preface
This document reports on the work of group 844 1 in the 8th semester. This report is
organized into two parts, main report and appendix. The main report contains seven
chapters. The first chapter starts as introduction and chapter 2 provides the overview of
problem modeling and problem analysis of the project is clearly outlined. Third chapter
is a review of requirements and specification. In fourth chapter, the design methodol-
ogy of the project using A3 model is described. In the chapter five, the algorithm of the
project is described. Finally chapter 6 and chapter 7, provides the implementation and
conclusion of the project. The second part of the report contains appendices with project
relevant, like linear predictive coding, speech and noise properties and about DSP proces-
sor TMS320C6713. This report ends with Matlab programs. All the associated material,
the Matlab code and C-programs along with project report are placed in the CD.
Aalborg,30th May 2005.

——————————-
Raju Muchanthula

1 Student of Applied Signal Processing and Implementation (ASPI).

Group 844: 30th May 2005 i

 Contents
1 Introduction 1

2 Problem Analysis 2
2.1 Problem Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.2 Speech Production . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.3 Process of Linear Predictive Coding . . . . . . . . . . . . . . . . . . . . 4
2.4 Noise Contaminated Signals . . . . . . . . . . . . . . . . . . . . . . . . 5

3 Requirements and specifications 7

3.1 Application requirements . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.2 Algorithm requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.3 System test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

4 System design 11
4.1 The A3 -model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
4.2 Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4.3 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4.4 Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.5 Test bed design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

5 Algorithm 17
5.1 Spectral Subtraction - Based on Minimum Statistics . . . . . . . . . . 17
5.2 Spectral Manipulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.3 Residual Manipulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

6 Implementation 24
6.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
6.2 DSP Architecture Mapping . . . . . . . . . . . . . . . . . . . . . . . . . 32

Group 844: 30th May 2005 ii

 CONTENTS

7 Conclusions 37
7.1 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
7.2 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

A Linear Predictive Coding 38

A.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
A.2 Linear Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
A.3 All Pole Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
A.4 Autocorrelation Method . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
A.5 Covariance Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
A.6 Levinson Durbin Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 42

B Speech and Noie 43

B.1 Speech . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
B.2 Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

C DSP 45
C.1 TMS320C6000 DSP platform . . . . . . . . . . . . . . . . . . . . . . . 45
C.2 Functional Units . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
C.3 Register File . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
C.4 TMS320C6713 DSP . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
C.5 C6713 Memory and Peripherals . . . . . . . . . . . . . . . . . . . . . . 47
C.6 TMS320C6713 DSP Starter Kit (DSK) . . . . . . . . . . . . . . . . . . . 47
C.7 Tools and Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
C.8 Algorithm Standard . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
C.9 Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

Bibliography 52

Group 844: 30th May 2005 iii

 Introduction
1
The Advances in telecommunications over the last few decades has dramatically affected
the way people live and communicate. Technological progress has made communication
systems reliable and affordable and mobile communication has now become omnipresent.
Resulting in the freedom and flexibility which introduces new challenges, one of the most
prominent is being the suppression of background noise. Mobile phones often used in
noisy environments, the result being additional unwanted noise signals are transmitted
with the speaker’s voice over the channel. At the receiving end the speaker’s voice can
become unintelligible amongst the noise. Researchers have been working for decades to
suppress background noise as much as possible for GSM instrument users and to posi-
tively influence the intelligibility of speech in such noisy environments. Suppression of
the background noise is important not only to improve the quality but also maintain the
intelligibility of speech [10].
The problems with noise are quite complex. Noise is composed of several different
sounds that are distinguishable from one another by temporal and spectral differences.
Noise components are classified into different types. Convolution noise, the unwanted
echo in hands free telephone system it can be called an acoustic noise or echo, which is
removed or reduced using adaptive filters. The other types of noises are additive noise
and noise due to non-linearities. Additive describes the nature of how a desired signal
and the noise energy are summed at the receiver input (there are impairment that are
multiplicative in nature).
Various methods were suggested overcoming the effect of noise on speech communica-
tions [10]. A single microphone scenario is used to analyze the speech data.
In microphone-based applications, which have a potential for incoming speech getting
corrupted by ambient noise captured by the microphone. It is especially suitable for
systems in which an acoustically isolated noise reference is not available, such as: Hands-
free Cell Phone Kits, Speakerphones, Intercoms, Teleconferencing Systems, Headsets,
As a front-end to a Speech Recognition system, Any microphone-based application that
needs to eliminate undesired noise.
The scope of the project is to reduce the background noise sufficiently from a given noisy
speech. To get the best noise reduction, an algorithm running in real time is chosen for
the processing. The algorithm is too implemented on the Texas Instruments DSP Kit,
TMS3206713 with suitable changes according to the architecture provided.

Group 844: 30th May 2005 1

 Problem Analysis
2
In this chapter, the production of speech will be introduced. Furthermore transmission
of noise free and noise contaminated signals will presented. This leads to an overview
of the system. The problem analysis consists of two sections. The first section concerned
with problem modeling. The later sections deal with speech generation, linear prediction,
solution to optimal predictor and transmission of noise contaminated signal.

2.1 Problem Modeling

The goal of this project is to develop an algorithm that can reduce the background noise to
facilitate conversation in noisy environment. The proposed system uses one microphone
for the input of the system. An illustration of the system can be seen in the figure 2.1.
The input signal to the microphone is noisy speech. In the filter, an estimation of the
background noise is subtracted from the signal picked by the microphone, so the output
is noise suppressed speech signal. In this project the major factor but maintaining the
perceptual quality and noise reduction are the main tasks.
In the following section, the speech generation and analysis are described before analyz-
ing the problem. The production of speech can be modeled with three blocks [3, Ch.2 and
3]; the generator, the vocal tract and the radiation. These blocks are depicted in figure 2.2.

2.2 Speech Production

Physical speech production originates in the lungs, where the airflow begins. The actual
sound is formed while air flows through the larynx and vocal tract. The larynx consists of
the cricoids cartilage, vocal folds and arytenoids cartilage. The vocal tract can be divided

Noise and Noise

Speech Suppression Speech

User Application

Figure 2.1: An illustration of one microphone scenario

Group 844: 30th May 2005 2

 2.2. SPEECH PRODUCTION

into three areas: oral pharynx, nasal cavity and mouth. The tongue, lower jaw and lips
have the most effect on the form of the vocal tract to produce distinct sounds.

Impulse train
Generator

Vocal tract Radiation

Model Model
Voiced/
Unvoiced S(n)
H(z) R(z) Speech
switch
White Noise
Generator

Figure 2.2: Block diagram of speech production.

Depending on the activity status of the vocal folds, the sounds can be roughly divided into
two extreme cases [6]. First, the vocal folds can open and close periodically, generating
in this way a train of pulses (glottal pulse). This gives the sound its voiced nature, i.e.
periodicity in time and harmonic structure in frequency. Second, the vocal folds can just
be open, with the airflow forming turbulence between the folds and the sound very noise-
like. Unvoiced sounds are formed this way. The excitation of the voiced sounds is usually
modeled by the train of pulses, whereas unvoiced sounds use random noise excitation.
The figure 2.2, is an illustration of the applicability of linear prediction for speech model-
ing [5]. The vocal tract model H(z) and lip radiation model R(z) are excited by a discrete
glottal excitation signal. For voiced speech the source excitation is an impulse train gen-
erator driving a glottal shaping filter and using local pitch period estimation. For unvoiced
speech the source excitation comes from a random noise generator.
Now, the knowledge about the generator can be used to analyze the speech. Production of
speech can be modeled as an AR-model, as the vocal tract can be modeled as an all-pole
filter [11]. Since the signal from the generator is assumed white, the inverse of the vocal
tract transfer function can be used to whiten the signals. The whitening filter can be found
using linear prediction, which is covered in detail later in this chapter and in appendix A.
For voiced sounds the glottis introduces two extra poles. It is desired to remove the
effect of these poles, in order to focus the parameters of the AR-model on the vocal tract.
The radiation can be modeled as a single-zero filter, which causes one of the poles to be
canceled. A preemphasis filter with the transfer function

R(z) = 1 − z0 · z−1 (2.1)

with 0.9 < z0 < 1 will cancel the other pole [3, p. 330]. This filter must be applied prior
to modeling the vocal tract.
The approach described here is model-based or parametric in contrast to a non-parametric
approach. The advantage of this approach is that it is possible to manipulate e.g. the
power spectrum or autocorrelation function without destroying the speech estimate since
the information about the pitch period is still maintained in the residual.

Group 844: 30th May 2005 3

 2.3. PROCESS OF LINEAR PREDICTIVE CODING

2.3 Process of Linear Predictive Coding

Linear predictive coding is a way of representation of speech which can be transmitted

efficiently through a digital channel. This method is based on linear predictive analysis.
The basis is the vocal tract which is modeled as all-pole filter. The purpose of LPC
in speech is shown in figure 2.3. The analysis results a set of LP coefficients and a
residual signal. Instead of transmitting the whole signal, the residual signal is encoded
and transmitted together with LP coefficients. At receiving end the speech is synthesized
by LP coefficients and residual signal.

Transmitter side Receiver side

Residual Estimated
Speech signal speech
Analysis filter Synthesis
s(n) filter ŝ(n)

Transmission
channel
LPC LPC
Coefficients Coefficients

Figure 2.3: Block diagram of Process of LPC

2.3.1 Linear Predictive Analysis

Linear prediction analysis of speech is one of the most important speech analysis tech-
nique. The assumption is that speech is short time or locally stationary for analysis. The
analysis is an extraction of prediction Coefficients, i.e an determination of formants cre-
ated by vocal tract. As shown in the figure 2.4 the speech is filtered with a prediction
error filter. The predictor is to predict the actual sample based on linear combination of
the past samples. If the predictor is optimal, the residue is spectrally white. The predictor
optimization is described in appendix A.

Linear Ŝ(n) −
S(n) ∑ e(n)
Speech predictor Residue
+

Figure 2.4: Block diagram of the prediction error filter.

2.3.2 LPC synthesis

Till now the focus was on analyzing the signal with the purpose of identifying the pa-
rameters of a system, satisfying AR constraint by minimizing the prediction error. If the
prediction error is ’white’, then the estimated signal is good fit, so the synthesized signals
having similar statistical properties as the original one. Then exciting the synthesis filter

Group 844: 30th May 2005 4

 2.4. NOISE CONTAMINATED SIGNALS

with the system function H(z) using a white noise signal, the filter’s output will have a
spectrum close to the original signal. The process is shown in the figure 2.5.

1 1
H(z) = = (2.2)
1 + ∑M
i=1 ai z
−i A(Z)
In the equation 2.2, ai are the LP coefficients found from the original signal. The transfer
function 1/A(z), G is a gain factor and it is excitated by the input signal e(n), and thereby
producing the speech signal s(n).

G
e(n) A(z) s(n)

Figure 2.5: The inverse system identification problem.

2.4 Noise Contaminated Signals

If the speech signal is corrupted by noise, the all-pole model no longer applied to approx-
imate the speech spectrum estimation closely and simply. Assume that the speech signal
is corrupted by an additive and uncorrelated background noise, then the signal model be-
come a pole-zero model, is also called as autoregressive moving average (ARMA) model.
In a system with pole-zero model the estimation of the parameters is inherently nonde-
terministic and nonlinear [6]. There are few solutions for pole-zero model like Gauss-
Newton method, but it is not guaranteed for the optimal solution [2, p.326]. So spectral
subtraction is considered as a solution to suppress the background noise.

2.4.1 Proposed Solution

Power Power Spectrum r̂(t)

Spectral IFFT LPC
spectrum manipulation
Px (ω) Subtraction Ps̃ (ω) Pŝ (ω) rŝ (k)

LPC

Speech Noisy Estimated

+ noise Analysis residual Residual Synthesis speech
filter manipulation filter
x[m] = s[m] + n[m] ẽ[m] ê[m] ŝ[m]

Figure 2.6: LPC-transmission using spectral subtraction

Spectral subtraction, a method to estimate the spectrum of a signal observed in additive,

uncorrelated noise. The estimate is obtained by subtracting an estimate of the noise spec-
trum from the noisy signal spectrum. The estimation of the noise is difficult because of

Group 844: 30th May 2005 5

 2.4. NOISE CONTAMINATED SIGNALS

the non stationarity of the process. Based on few assumptions speech is considered as
locally stationary for analysis [12, p. 208]. The average noise power is approximately
the same prior to speech activity as during speech activity. With these assumptions, the
estimate of the signal spectrum Pŝ (ω) is obtained by [9, p.114]:

Pŝ (ω) = Px (ω) − Pn (ω), (2.3)

where Px (ω) is the spectrum of the noise contaminated signal and Pn (ω) is the noise
spectrum estimated during no speech activity. It is difficult to remove entire noise by this
method, as the assumptions are not entirely correct. Some of the noise components are
not additive resulting in a correlation between the speech and noise in someway. Due
to these reasons, the performance of the method depends on the estimation of the noise
spectrum [9].
The classical spectral subtraction method does not deal with signal phase information.
Same noisy signal phase is used to reconstruct the estimate of the speech signal. Even the
manipulation of the magnitude spectrum distorts the pitch period. Thus, another model is
chosed which is based on linear prediction. This model uses the envelope of the amplitude
spectrum without considering the phase information. The proposed method i.e., spectral
subtraction using LPC is shown in figure 2.6.

Group 844: 30th May 2005 6

 Requirements and
specifications 3
As mentioned in the problem analysis, chapter 2 the project deals with the noise sup-
pression. So the requirements for the development and implementation of the algorithm
are mentioned in this chapter. With these desired requirements, the specifications for the
algorithm implementation are discussed.

3.1 Application requirements

The system is only considered as a noise suppression device, which gives the two main
objectives:

• The SNR of the signal should be improved.

• Intelligibility /perceptual quality should be maintained.

The boundaries specified for SNR are between 0 dB and 35 dB, as mentioned in [3,
p.586]. The limit is considered so as the sysytem can function efficiently.

3.2 Algorithm requirements

The system described in the Problem Analysis utilizes a model based description of the
speech production. It is thereby possible to set up specific requirements to the parameters
in this model. With starting point in figure 2.6, requirements to the different parameters
will now be presented.

3.2.1 Frame characteristics

The speech is assured to be quasi-stationary resulting in the specification of the type of

window and its length, amount of overlap between frames.

• Frame Size: The frame size had been set to 25ms, if the analysis frame shorter than
20ms results in roughness, while increasing the frame size decreases the musical

Group 844: 30th May 2005 7

 3.2. ALGORITHM REQUIREMENTS

overlap

frame
shift/burst

(a) Illustration of frames, burst and overlap.

Frame, sent to the frame based execution part.

Corresponding burst, used for calculation of the
residual using the calculated coefficients for the
corresponding frame.
Adjustable delay
frames

bursts

(b) Illustration of the adjustable delay between frames and

bursts. As seen, the bursts form a time continuous signal.

Figure 3.1: The relation between frames, bursts and overlap.

noise considerably. If the frame size is too long, it results slurring [12]. The frame
and burst are shown in the figure 3.1.
• Window Selection: The window determines the portion of the speech signal that
is to be processed by zeroing out the signal outside the region of interest. The
ideal window frequency response has a very narrow main lobe which increases
the resolution, and decreases the side lobes (or frequency leakage). Since such
a window is not possible in practice, a compromise is usually selected for each
specific application.
There are many possible windows, such as rectangular, hanning, hamming window.
The rectangular window has the highest frequency resolution due the narrowest
main lobe, but has the largest frequency leakage. Due to the high frequency leakage
produced by the larger side lobes, rectangular windowed speech looks more noisy
than hamming or hanning windowed speech. This undesirable high frequency leak-
age between adjacent harmonics tends to offset the benefits of the flat time domain
response of the rectangular window. As a result, rectangular windows are not usu-
ally used for speech spectral analysis. On the other hand, the trapezoidal windows
such as hamming and hanning window has smaller frequency leakage but lower
resolution. So, they produce a smoother spectrum than the rectangular window [4].
Since LP model is determined on frame-by-frame basis, original signal should be
segmented into frames. The segmented signal is also required to be windowed

Group 844: 30th May 2005 8

 3.2. ALGORITHM REQUIREMENTS

frame by frame. A properly windowed frame is easier to analyse in frequency do-

main. Additionally, frequency leakage of main lobe is reduced by trapezoidal win-
dow, such as hamming window or hanning windows. The window size is selected
as length of the frame. Determination of window type should be done experimen-
tally by doing Matlab simulations with different types of windows.
• Frame overlap: When sliding the window through the signal, the overlap is nec-
essary to prevent discontinuities at frame boundaries. The amount of overlap is
usually taken to be 20% of the frame size. For large frames, 10% may be excessive
and might causes slurring of the signal. So amount of overlap will be determined
from experimental results [12].

3.2.2 Analysis filter

The purpose of the analysis filter is to analyze the signal and identifying the parameters
of a system, satisfying AR constraint by minimizing the prediction error. If the prediction
error is ’white’, then the estimated signal is good fit, so the synthesize signals having
similar statistical properties as the original one. The objective of the analysis filter is to
cancel the characteristics of the vocal tract in the observed signal and by it reproducing an
estimate of the excitation signal at the output of the filter. This can also be considered as
canceling the formants in the observed signal, which is the same as minimizing the error
signal. In order to do this, it is necessary to have enough coefficients in the analysis filter
to model the formants. In [3, fig. 5.7c] it is seen, that with a sampling frequency of 8
kHz, the residual energy is not further minimized when the prediction order exceeds 10.
Furthermore, an algorithm for calculating the LP coefficients should be suited for real-
time implementation. A natural choice would therefore be the Levinson-Durbin algo-
rithm, which also applies to the LPC-block.

3.2.3 Manipulation of residual

In the case of a noise contaminated speech signal as the observed signal, the AR-model
analysis filter will have a noisy residual as an output. Manipulation of this noisy residual
should lead to an emphasis of the voiced segments i.e. the impulses in the residual.

3.2.4 FFT order

The length of the FFT must be long enough to give a sufficient frequency resolution. The
preffered length is 256 samples. The minimum FFT order corresponding to a given frame
size is adequate [12].

3.2.5 Power Spectral Subtraction

This block is the main SNR improving part of the system and the ideal goal is to entirely
remove the noise spectrum from each analyzed frame, which is not possible. The output

Group 844: 30th May 2005 9

 3.3. SYSTEM TEST

should not be a time estimate of the speech signal, which is the classical usage of spectral
subtraction, but instead an estimated power spectrum of the speech signal to be used in
the IFFT-block in order to obtain the autocorrelation. From the autocorrelation, the LP
coefficients are calculated.
The block should be able to estimate the noise spectrum under different types of noise
and varying noise levels.

3.2.6 Manipulation of power spectrum

Since it is not possible to entirely remove the noise from the observed signal, manipula-
tion of the estimated speech power spectrum is necessary. The results of this manipulation
should be an estimated speech power spectrum in which the randomly varying parts due
to noise is further reduced.

3.2.7 Synthesis filter

After spectral subtraction, power spectrum manipulation, and autocorrelation (ifft), a es-
timated time wave of speech is reconstructed using the manipulated LPC coefficents and
manipulated noisy residual. Then the synthesis filter’s output will have a PSD close to
the original signal as long as the prediction order 10 is adequate.

3.3 System test

The purpose of the system test is to test the system as a black box. There will be per-
formed a test of the individual parts of the system.
The system test will be carried out with a noise contaminated speech signal. The test will
show whether the requirements are maintained, with respect to SNR, intelligibility.

Group 844: 30th May 2005 10

 System design
4
This chapter describes the system design. The design is based on Rugby model. In relation
with rugby model the A3 model, Application, Algorithm and Architecture are discussed.
The application is described for real time implementation. Algorithm is divided into
two parts, burst based and sample based. The architecture is Texas Instruments DSP,
TMS320C6713 DSK [13].

4.1 The A3 -model

The Rugby model is a conceptual frame work, in which designs, design processes are
expressed in order to analyze the problem. This model is used to evaluate the design
covers all the domains at the various abstraction levels. Rugby model has four domains
for design process, namely Computation, Communication, Data, and Time.
As shown in the figure 4.1 rugby model has starts with an idea or project proposal (higher
abstraction level) and goes through four domains with different abstraction levels to de-
velop the (lower abstraction level) final system. All abstraction levels are treated these
domains resulting in the final design. In this project the system design is based on the
A3 -model as shown in the figure 4.2. The model takes into consideration that the algo-
rithm may be modified before moving to the architecture domain. Furthermore the move
from the algorithm domain to the architecture domain i.e the feed back from the fixed
architecture can be performed several times.

Time

Computation
Final
Idea System

Communication

High level Data Low level

Abstraction Abstraction

Application Algorithm Architecture

Project time line

Figure 4.1: The Rugby meta model in relation with the A3 -model.

Group 844: 30th May 2005 11

 4.2. APPLICATION

Application
1

Algorithm 2

Architecture

Figure 4.2: The A3 -model.

4.2 Application

The application of the project is to develop a noise suppression algorithm, which is must
be an adaptive for give architecture. The system black box is illustrated in the figure 4.3.
While considering the timing constraints, the system is expected to run as a real time sys-
tem. For that latency must not exceeds 30ms, in order to the audio or visual requirement.
The input is a noise contaminated speech signal, x[m], and the output is an estimate of the
noise free speech signal, ŝ[m].

Noise contaminated Estimated

speech signal Noise speech signal
X[m]=S[m]+n[m] Suppression ^
S [m]
System

Figure 4.3: The overall system shown as a black box.

4.3 Algorithm

The computational complexity and desired output are considered as the most important
factors for choosing the algorithm. Based on the problem analysis and the application
description, the system is partitioned into processes. The processes are groupings of
functionalities, which are defined in the problem analysis.
The system consists of different parts, that will be executed with different inputs. One
part of the system is running as sample based execution, while the other two parts run
as burst based and frame based execution. In the sample based part, circular buffers are
available, providing the possibility of the frame overlap. The overview of the algorithm
is shown in the figure 4.5. These different processes of the algorithm to be analyzed step
by step. The processes and the interface between these are described in short terms in
table 4.1.
The next step in the design process is a partitioning of the processes into functions. It is
defins which processes are concurrent and therefore can be executed in parallel. It also
describes how parameters are be passed between the functions.

Group 844: 30th May 2005 12

 Process Input Output Description
Preemphasis filter Speech samples, x[m] Blocks of speech samples, either Accounts from the effect of the lip/radiation
frame or burst. from the mouth. The outputs are related as
depicted in figure 3.1(b)
LPC Block of data, frame LP coefficients, a Calculates the LP coefficients for the analysis
filter.
Analysis filter Block of data, burst, and LPC- Noisy residual, ẽ[m] Whitens the speech signal and creates the
coefficients, a residual using prediction error filtering.
Residual manipulation Noisy residual, ẽ[m] Estimated residual, ê[m] Suppresses noise in the residual signal.
Group 844: 30th May 2005

Synthesis filter Estimated residual, ê[m] and LP Estimated speech samples, ŝb [m] Synthesizes the speech signal from the resid-
coefficients,ˆa ual and the LPC-coefficients.
Power spectrum Frame of data, frame Power spectrum, P x (w) Calculates the power spectrum of the speech
frame.
Spectral subtraction Power spectrum, P x (w) Noise suppressed power spec- Suppresses noise in the power spectrum.
trum, Ps̃ (w)
Spectrum manipulation Noise suppressed power spec- Estimated speech power spec- Estimates the power spectrum using a model-
trum, Ps̃ (w) trum, Pŝ (w) based approach.
IFFT Estimated speech power spec- Autocorrelation function, rŝ (τ) Calculates the autocorrelation from the power

4.3. ALGORITHM
trum, Pŝ (w) spectrum.
d
LPC Autocorrelation function, rŝ (τ) LP coefficients, â Calculates the LP coefficients from the auto-
correlation function.
Deemphasis filter Estimated burst of speech sam- Estimated speech samples, ŝ[m] The inverse of the preemphasis filter.
13

ples, ŝb [m]

Table 4.1: Description of the processes and their interfaces.

 4.4. ARCHITECTURE

Requirements

Specifications

Matlab model

C−model

TMS320C6713 DSK

Figure 4.4: Transformation in the System Design

4.4 Architecture

When mapping the algorithm to the architecture, several aspects need to be considered,
like how to implement the chosen in an optimal manner for given architecture on different
abstraction levels. The level of implementations are shown in the figure 4.4.
The basic algorithm implemention and testing to be done in MATLAB. Then the matlab
code is converted into C. Considering the architecture the C and Matlab functions could
be optimized with minimum changes. To facilitate functional debugging and easy to
access to the internal variables of the algorithm. This should be ensure compatibility with
the C-compiler included in CCS, allowing the code to be transferred with alternations.
In the next section a test bed will be designed, so that there is a proper way of testing the
system.

4.5 Test bed design

To be able to test the algorithms, and thereby the system, a testbed design environment
must be designed.
The design of the testbed is based on stepwise refinement, so that the system at the highest
level of abstraction is tested first. A single process or function will be tested at a time,
keeping the number of adjustable parameters to a minimum. It must be emphasized,
that each process and function will have to fulfill the requirements in the requirements
specification.

Group 844: 30th May 2005 14

 Frame based execution

Power Px (w) Ps~(w) Spectrum P~(w)

s rs (τ)
Spectral IFFT
spectrum substraction manipulation LPC
Figure 4.5: The black box of the system.

Windowed frame ^a
Group 844: 30th May 2005

LPC

Windowing
a
Burst based execution

Noisy Estimated
Burst residual Residual residual Synthesis Deemphasis
Analysis filter ~e[m] manipulation filter filter
^~
e[m] S^b [m]

Framming
Sample based execution

4.5. TEST BED DESIGN

Circular buffer Circular buffer

Estimated
Speech Preemphasis speech
+ filter ^S[m]
noise x[m] X[m] ^
S[m]
in_buf out_buf
15
 4.5. TEST BED DESIGN

First a Matlab model will be developed, so that a fully functioning mathematical model
can be tested. The Matlab model gives the possibility to implement one module at a time,
replacing the Matlab functions with C-functions and test the C-functions in connection
with the already working Matlab-functions. Thereby, only one function is changed at a
time, making it possible to track errors down to a single function.
The processes and their functions will be tested on the fixed architecture. This will be
performed by writing some test-data in memory, and then running the function. The test
data will be the same data that is used when testing the Matlab functions.
After the testbed has been designed, the design of each process in the algorithmic and ar-
chitecture domain will be performed. The test of the process will performed in connection
with this design.

Group 844: 30th May 2005 16

 Algorithm
5
The purpose of this chapter is to introduce the algorithms that will model and analyze
the performance. As metioned in system design this chapter is a division of the algorithm
in to number of blocks, named spectral subtraction, spectral manipulation and residual
mainipulations. For more related detail information on implemetation see chapter 6.

5.1 Spectral Subtraction - Based on Minimum Statistics

The performance of speech signals degrades in the presence of noise. To reduce the
degradation, a model is proposed to represent the spectral changes of speech signal uttered
in noisy environments. The spectral subtraction method is a well-known noise reduction
technique. The standard algorithms for spectral subtraction usually need a voice activity
detector (VAD), such that the noise spectrum can be estimated during non-speech activity.
But VAD’s performance is often degraded considerably under noise conditions. Most
implementations and variations of the basic technique advocate subtraction of the noise
spectrum estimate over the entire speech spectrum. In general noise is mostly colored
and does not affect the speech signal uniformly over the entire spectrum. This chapter
describes an alternative robust algorithm, which make use of minimum statistics while
estimating the noise spectrum. The described algorithm is based on the article [8].

5.1.1 Overview

Spectral subtraction is a noise power estimator and subtraction rule which translates the
SNR into a spectral weighting factor, such that low SNRs are attenuated and high SNRs
are not modified. Spectral subtraction method is computationally efficient.
The assumption is that the power spectrum of signal is corrupted by uncorrelated noise,
is equal to the sum of a speech spectrum and noise spectrum.

5.1.2 Terminology

Below are some of the more common terms that are encountered when discussing the
spectral subtraction algorithm.

• Spectral floor: The spectrum components of the processed signal below a certain
lower bound is called spectral factor.
• Musical noise: The noise which is generated in the spectral subtraction process.

Group 844: 30th May 2005 17

 5.1. SPECTRAL SUBTRACTION - BASED ON MINIMUM STATISTICS

• Short time power: The estimation of the power using the short-time.
• Broadband noise: In the frequency domain, a broadband noise has a continuous
spectrum, that is present at all frequencies in a given range. This type of sound
often referred to as noise because it usually lacks a discernible pitch.
• Comb filter: The filter has multiple pass bands and stop bands. It has a frequency
response with a periodic function of w and a period 2π/L, where L is a positive
integer.

5.1.3 Basic idea

In the spectral subtraction, the minimum of subband noise power with a finite window to
estimate the noise floor. A periodogram of Px (λ, k) is shown in the figure 5.1, the short
time power estimate of noisy speech signal shows distinct peaks and valleys.

−3
x 10
16 Px
P
min
P
n
14

12

10
Power [W/Hz]

200 400 600 800 1000 1200 1400

λ [n ⋅ L/2 ~ 10ms]

Figure 5.1: Periodogram, Px (λ, k), for fs = 8kHz (NFFT = 256, k = 25) of the input signal
.

The basic idea behind the algorithm is an utilization of these peaks and valleys. The
peaks are assumed to speech activity and the valleys are used to obtain the estimation
of subband noise power. To obtain the noise power estimates the data window for the
minimum search must be large enough to bridge peaks of speech activity.

5.1.4 Algorithm overview

A block diagram describing the implemented algorithm for spectral subtraction is shown
in figure 5.2.

Group 844: 30th May 2005 18

 5.1. SPECTRAL SUBTRACTION - BASED ON MINIMUM STATISTICS

 
x[m] = s[m] + n[m] xw [m] |X(λ, k)| Ŝ(λ, k)
 

Windowing FFT ×
Q(λ, k)
kosub
Pn (λ, k)

Noise Computation
power of spectral
estimation Pn (λ, k) weighting

Figure 5.2: Block-diagram the spectral subtraction algorithm.

The signal x[k] is assumed that the sampled and windowed signal, x[k] is sum of a zero
mean speech signal s[k] and a zero mean noise signal n[k] is:

x[k] = s[k] + n[k] (5.1)

Further assumption is that, speech and noise are statically independent. The data window
w[k] and the short-time fourier transform of x[k] is a given by,

xw [λ, k] = x[λ, k]w[k] for k = 1, . . . , Lwindow (5.2)

where λ indicates the time index, k indicates the frequency bin index and the discrete
frequencies.Ωk = N2πk
FFT
, k = 0, 1, . . . , NFFT − 1. By computing the FFT of individual win-
dowed frames of length NFFT . Then the short-time power spectrum from the equation 5.1
is derived as:

Px (λ, k) = Ps (λ, k) + Pn (λ, k) (5.3)

where Px (λ, k), Ps (λ, k) and Pn (λ, k) are the short time power spectrums of the noise con-
taminated speech, the speech and the noise signal respectively.

5.1.5 Subtraction Rule

To obtain |X(λ, k)|2 , the smoothed estimate of short time power of x w [k] with a first order
recursive network is given by,

|X(λ, k)|2 = γ|X(λ − 1, k)|2 + (1 − γ)|X(λ, k)|2 (5.4)

Where γ is smoothing constant (γ . 0.9). Now the aim is to estimate the short time
 
magnitude spectrum of the clean speech signal, Ŝ(λ, k). This can be done by using
following subtraction rule:

Group 844: 30th May 2005 19

 5.1. SPECTRAL SUBTRACTION - BASED ON MINIMUM STATISTICS

(
  ksubf · Pn (λ, k) if |X(λ, k)| · Q(λ, k) ≤ ksubf · Pn (λ, k)
Ŝ(λ, k) = (5.5)
|X(λ, k)| · Q(λ, k) otherwise

ksubf is a spectral floor constant. It prevent the spectral components of Px (λ, k) from
descending below the lower bound ksubf · Pn (λ, k). The variable gain factor Q(λ, k) is
given by
s
Pn (λ, k)
Q(λ, k) = 1 − kosub (λ, k) (5.6)
|X(λ, k)|2

where kosub (λ, k) is an oversubtraction factor, which is a function of the signal-to-noise

ratio, SNR.

5.1.6 Short time noise power estimation

An estimation of the short time noise power Pn (λ, k) is done from the short time power
Px (λ, k), which is estimated by filtering the instantaneous short time power |X(λ, k)| 2 , to
obtain a smoothed estimate i.e.,

Px (λ, k) = αPx (λ − 1, k) + (1 − α) |X(λ, k)|2 (5.7)

where 0.9 ≤ α ≤ 0, 95 is a smoothing constant. The smoothing with a fixed parameter α

widens the peaks of speech activity of the smoothed estimate Ps (, k). This leads to inac-
curate noise estimate. Using (5.7) Pn (λ, k) is estimated by finding a weighted minimum
Px,min (λ, k) within the last D frames of Px (λ, k). The estimate is given by

Pn (λ, k) = komin Px,min (λ, k) (5.8)

komin is a factor to compensate the bias of the minimum estimate. The minimum is ob-
tained as a comparison between a minimum, calculated every M frames, of the last D
frames of Px (λ, k) given by

PDmin (λi , k) = min (Px (λ j , k)) j = i − D + 1, . . . , i (5.9)

and the actual value of PX (λ, k). The minimum is obtained as the minimum of the last W
minimums, that is D = W · M. By splitting up D the minimum is obtained after only M
frames instead of after D frames.

5.1.7 Estimation of SNR

To control the oversubtraction factor k osub (λ, k), the SNR in each subband is computed. If
the SNR is high, the oversubtraction level is low and vice versa. An estimate of the SNR
is:  
Px (λ, k) − min(Pn (λ, k), Px (λ, k))
SNR(λ, k) = 10 log (5.10)
Pn (λ, k)

Group 844: 30th May 2005 20

 5.2. SPECTRAL MANIPULATION

K osub

5
4
3
2
1

−10 −5 0 10 20 30 40 SNR (dB)

Figure 5.3: Graph of the subtraction factor kosub Vs SNR

Large values of kosub leads to speech distortion. As shown in the figure 5.3, the optimal
value of kosub is found for best noise reduction with less amount of musical noise is
smaller for higher SNR. For kosub = 1, the SNR ≥ 35 dB and as increasing the k osub value
the SNR will be SNR ≤ −5 dB.
The SNR is calculated for each short-time frame, so the k osub value also varies for frame
to frame. The actual value of kosub used in equation 5.6 is given by:

kosub = k0 − (SNR)/s, −5 ≤ SNR ≤ 35, (5.11)

where kosub is value of k 0 at SNR = 0 and 1/s is slope of the line.

5.2 Spectral Manipulation

After spectral subtraction, the estimated speech power spectrum still contains noise. This
noise can be suppressed by spectral manipulations. The manipulation can be done in two
ways:

• Spectral smoothing
• Floor setting

5.2.1 Spectral Smoothing

The noise components, which abruptly comes up for a short time in the power spectrum
are observed as noise. This type of noise can be reduced by spectral smoothing in time
domain. For smoothing, the power spectrum is filtered with a lowpass filter. The designed
lowpass filter should be able to suppress the short time varying components or noise
components in the power spectrum without disturbing the formant frequencies and short
time stationarity. The smoothed spectrum should be able to produce noise suppressed lPC

Group 844: 30th May 2005 21

 5.3. RESIDUAL MANIPULATION

coefficients for synthesis. The smoothing in the time domains is capable of changing or
modifying the amplitudes of the spectrum. This process can be done by filtering.
If the estimated speech spectrum contains rapidly or slowly varying frequency compo-
nents compared to original speech spectrum, in such cases it is difficult to smooth the
spectrum. This effect can be controlled by spectral smoothing in frequency domain. So
the estimated speech spectrum is needed to be filtered with a lowpass filter.
By filtering the spectrum, it should capable to model the envelope of the estimated speech
spectrum. In the filtering process, it is needed to maintain the frequency components
should not be delayed. The same formats should retained even after smoothing.

40

35

30

25

20

15
r

10

−5

−10
−400 −300 −200 −100 0 100 200 300 400
Lag [samples]

Figure 5.4: Autocorrelation of noisy residual signal

5.2.2 Spectral Floor Setting

In the spectral subtraction method, subtracting an estimate of the noise power spectrum
from the noisy speech power spectrum results in estimated speech power spectrum and
setting the negative differences to zero [12].

5.3 Residual Manipulation

As described in system design, the algorithm is divided in two parts. Till now, the discus-
sion and description was about first part, spectral subtraction and spectral manipulation
with frame based execution. Now, the residual manipulation using burst based execution.
The ultimate challenge is trying maximum possible ways to suppress the background
broadband noise and increase the SNR in which residual manipulation also plays major
role.
The corrupted harmonics in noisy residue are regenerated by filtering with a comb filter.
But the comb filter needs a pitch detector for detecting the pitch period in the residue.

Group 844: 30th May 2005 22

 5.3. RESIDUAL MANIPULATION

−50

−60

−70

Power/frequency (dB/Hz) −80

−90

−100

−110

−120
0 1 2 3 4 5 6 7
Frequency (kHz)

Figure 5.5: Noisy residual signal in frequency domain

This leads to a problem as sometimes it detects wrong pitch period and even it is quite
complex to construct a comb filter to suppress the noise level in the residue. So in order
to avoid the pitch detection, another method called autocorrelation is considered to reach
the desired response.
The autocorrelation of a noisy residual signal is shown in the figure 5.4. The duration
between the impulses is considered as pitch period. And the corresponding power spec-
trum is shown in the figure 5.5. The power spectrum clearly shows that the frequency
components are amplified. Now, by filtering the spectrum, the noise level of the residual
signal is reduced.

Group 844: 30th May 2005 23

 Implementation
6
The purpose of this chapter is to describe the implementation of algorithm as mentioned
in chapter 4 that we will analyze the performance. The purpose and functionality of
each individual block of Matlab model is described. The implementation is performed on
different development levels: First the basic model is in Matlab, C in Matlab and C in
6713DSK.

6.1 Overview

As discussed in design chapter to implement the chosen algorithm in an optimal manner.

In this chapter the criteria is to fit the algorithm for a fixed architecture of 6713DSK. The
system is tested for both functionality and performance.

6.1.1 Input Signal model

Different types of signals are considered for implementing the algorithm.

• Speech signal without noise.

• Speech signal with an additive uncorrelated noise.

The clean speech signal of male speaker who speaks a sentence “watch the log float in
the wide river”. The additive noise signal is a pink noise. The figure 6.1 shows noise free
signal and noisy speech signal.

6.1.2 Circular buffer

The circular buffers are used in the algorithm for real-time implementation. Digital signal
processors which support circular buffers automatically generate and increment pointers
for memory accesses which wrap to the beginning of the buffer when its end is reached,
thus saving the time and instructions otherwise needed to ensure that the address pointer
stays within the boundary of the buffer, and speeding the execution of repetitive DSP
algorithms [14].

Group 844: 30th May 2005 24

 6.1. OVERVIEW

0.5 0.5
Pink noise
Speech
0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0

−0.1 −0.1

−0.2 −0.2

−0.3 −0.3

−0.4 −0.4

−0.5 −0.5
0 2 4 6 8 10 12 0 2 4 6 8 10 12
4 4
x 10 x 10

(a) Speech signal (b) Speech signal with pink noise

0.2 0.2 1
Hanning window

0.15 0.15 0.9

0.8
0.1 0.1

0.7
0.05 0.05

0.6
0 0
0.5
−0.05 −0.05
0.4

−0.1 −0.1
0.3

−0.15 −0.15
0.2

−0.2 −0.2 0.1

−0.25 −0.25 0
0 20 40 60 80 100 120 140 160 0 20 40 60 80 100 120 140 160

(c) A frame of a speech signal (d) Windowing a frame with a hanning window

0.15 0.2

0.15
0.1

0.1

0.05
0.05

0
0

−0.05
−0.05

−0.1
−0.1

−0.15

−0.15
−0.2

−0.2 −0.25
0 20 40 60 80 100 120 140 160 0 20 40 60 80 100 120 140 160

(e) A windowed frame of a speech signal (f) A burst of a speech signal

Figure 6.1: Speech Analysis

Group 844: 30th May 2005 25

 6.1. OVERVIEW

6.1.3 Frame, Burst and Windowing

The input signal of the system is a vector. The vector is made into frames with a length
of 160 samples and 20 percent overlap between two consecutive frames. A frame of the
input signal is shown in the figure 6.1.
The frame and burst relation is shown in the figure 3.1. 20 percent overlap of the frame
is used as a delay for the analysis. The same delay is maintained for the synthesis.
Each frame is windowed using hanning window. A windowed frame with well shaped
boundaries is shown in the figure 6.1.

6.1.4 Pre-emphasis filter

The pre-emphasis filter boosts the harmonic components of the input signal in order to
get a better estimation. The purpose of the pre-emphasis filter is shown in the figure 6.2.

Before filter
5
After filter
4.5

4

Px (λ = 100, f ) ×10−4

3.5


2.5

1.5
PSfrag replacements
1

0.5

0 500 1000 1500 2000 2500 3000 3500

f

Figure 6.2: Functionality of the preemphasis filter.

6.1.5 LPC Analysis Filter

In linear prediction, the speech waveform is represented by a set of parameters of an

all-pole model, called the linear predictive coefficients (LPC), which are closely related
to speech production transfer function. The LPC analysis essentially attempts to find
an optimal fit to the envelope of the speech spectrum from a given sequence of speech
samples. The analysis filter gives a residue which is used as an excitation signal for
synthesis.
In this algorithm, two LPC analysis filters with order 10 of a speech spectrum is shown in
the block diagram 4.5. The first LPC analysis filter in figure 4.5 residue and the second
LPC analysis filter coefficients are used for synthesis. Even the performance tested with
LPC coefficients of second LPC filter and residue of the same filter is used for synthesis,

Group 844: 30th May 2005 26

 6.1. OVERVIEW

but gave poor results. The linear prediction to the signal spectrum with a 11-pole and
sampled at 8 kz and residual signals are shown in figure 6.3. The signal power spectrum
compared with residual power spectrum and auto-correlation of the residue are shown
in 6.3.

0.2
−20

0.15
−25

0.1

−30
PSD[dB]

0.05

−35
0

−40
−0.05

−45 −0.1

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

−0.15
Normalized Frequency (×π rad/sample) 40 50 60 70 80 90 100 110 120

(a) Linear prediction to the signal spectrum with a (b) A residual signal from the 25th burst
11-pole and the signal sampled at 8kz

−4 −6
x 10 x 10
10 8
Pe
Px

6
5

5
Power [W/Hz]
r

0 3

−5

0
−60 −40 −20 0 20 40 60 0 500 1000 1500 2000 2500 3000
Lag [samples] f

(c) Auto correlation of the residual signal from a (d) Signal power spectrum vs residual power spec-
35th burst trum

Figure 6.3: Linear prediction analysis and Residual siganl

6.1.6 Spectral Subtraction

The spectral subtraction implementation is based on proposed algorithm 5. The power

spectrum of each individual frame is calculated with the FFT length of 256. The FFT
length should be greater than or equal to window length for the sufficient frequency res-
olution.

Group 844: 30th May 2005 27

 6.1. OVERVIEW

The main objective of the power spectral subtraction is to estimate the noise spectrum,
Pn (λ, k) and subtract it from the input signal power spectrum, Px (λ, k) to obtain the esti-
mate of speech power spectrum, Pŝ (λ, k). The process should be able to track different
types and levels of noise spectrum is shown in figure 6.4. The implemention of algo-
rithm is described as flow chart is shown in figure 6.9. For this different parameters are
considered while implementing.

Parameter Symbol Ranage

Bias compensation factor Omin 1 ≤ Omin ≤ 1.5
Spectral floor constant ksub f 0.01 ≤ ksub f ≤ 0.05.
Smoothing constant α 0.9 1 ≤ α ≤ 0.95
Amplitude spectrum smoothing γ γ ≤ 0.9
constant

Table 6.1: Description of parameter range.

The performance evalvation of the algorithm is determined by certain parameters are

shown in the table 6.1. These values are chosen on the basis of [Martin,1994].

6.1.7 Spectral Manipulation

After the subtraction, as mentioned in algorithm, the spectral subtraction, spectral ma-
nipulation is necessary as mentioned in chapter 5. The manipulation is done in three
steps. First step the envelope estimation, which estimates the envelope of the smoothed
spectrum inorder to describes the LPC model. Different types of filters are designed for
envelope estimation and smoothing. The envelope estimation a kind of low-pass filtering
is shown in 6.4.
The first order recursive filter is used for smoothing in frequency domain. The smooth-
ing in frequency domain is carried out to suppress the frequency components which are
introduced in envelope estimation is shown in 6.4. The next approach is smoothing in
time domain is to smoothing the Periodogram is shown in figure 6.4, is similar to that of
frequency smoothing.

6.1.8 IFFT

From the manipulated estimate of the speech power spectrum, the IFFT should calculated
the autocorrelation for the signal frame. In figure 6.5, the calculated autocorrelation for
the frame index λ = 95 is shown.

Group 844: 30th May 2005 28

 6.1. OVERVIEW

−4
x 10
−6 Averaged noise power spectrum
x 10 14 P
3 x
Pmin
P
n
12
2.5

10

Power [W/Hz]
8

1.5
6

1 4

2
0.5

100 200 300 400 500 600 700 800 900

0
0 1000 2000 3000 4000 5000 6000 7000 8000 λ [n ⋅ L/2 ~ 10ms]

(a) Averaged estimated noise spectrum (b) Estimated noise floor of noisy speech signal
( f = 8khz, NFFT = 256))

−5 −5
x 10 x 10
6 1.6
envelope Smoothed
Original

1.4
5

1.2

4
1
Power[w/hZ]

power[w/hZ]

3 0.8

0.6
2

0.4

1
0.2

0 0
4000 4500 5000 5500 6000 6500 7000 7500 8000 0 500 1000 1500
f[hZ] f[hZ]

(c) Envelope estimation of estimated signal power (d) Smoothed estimated power spectrum
spectrum

−5
x 10
−110 5

4.5

−120
4

3.5
Power Spectral Density (dB/Hz)

−130
Ps(λ=97, k)[w/khz]

−140 2.5

−150
1.5

1
−160

0.5

−170 0
0 1000 2000 3000 4000 5000 6000 7000 0 500 1000 1500 2000 2500 3000 3500 4000
Frequency (Hz) f[khZ]

(e) Periodogram of estimated power spectrum (f) Estimated speech power spectrum after spectral
manipulation.
Group 844: 30th May 2005 29
Figure 6.4: Results of Spectral subtraction and spectral manipulation
 6.1. OVERVIEW

10


4

rŝ (λ = 95, τ) 10−6


2

−2

−4

PSfrag replacements
−6

−100 −50 0 50 100

τ

Figure 6.5: Autocorrelation of an estimated speech frame calculated from the manipulated speech
power spectrum estimate.

6.1.9 Residual Manipulation

The residual signal is result of the LPC analysis filter. The residual signal would be zero,
as the prediction filter would exactly match all the poles there in. This phenomenon does
not occur often, so a 10th-order LPC analysis filter is used in this algorithm implementa-
tion. The residue of analysis filter of an order of 10 is shown in the figure 6.1.
Thus, the AR-modeling is not sufficient if the input signal is noise contaminated. So
noise suppression is necessary in order to get better synthesized speech. The burst based
execution is carried out to reduce the complexity in the process of noisy residue.
As proposed in algorithm, the autocorrelation method is considered to find the pitch pe-
riod in the residue. and corresponding power spectrum is calculated. Now a first order
lowpass filter is used to filter the spectrum. The filter responses is shown in the figure 6.6.

6.1.10 Synthesis filter

After finding the LPC coefficients from the autocorrelation, the estimated speech wave
form is reconstructed using the results of analysis filter, the manipulated LPC coefficients
and manipulated residual signal. It is called inverse system identification. The results
of the of the synthesis filter is shown reffig:estimate. The spectrograms of clean speech
signal, noisy spech signal and estimated speech signal are shown in the figure 6.7.

6.1.11 ANSI-C in MATLAB

The software is basically developed as a Matlab functions in ANSI-C. This allows easier
access to the internal variables, and makes it possible to verify the results of the individual
functions. The conversion of the software is based on simulations in Matlab. In the ANSI-

Group 844: 30th May 2005 30

 6.1. OVERVIEW

40 −55

35 −60

30
−65

25

Power Spectral Density (dB/Hz)

−70

20
−75
15
r

−80
10

−85
5

−90
0

−5 −95

−10 −100
−400 −300 −200 −100 0 100 200 300 400 0 1000 2000 3000 4000 5000 6000 7000
Lag [samples] Frequency (Hz)

(a) Auto correlation of the residual signal (b) Periodogram of residual Signal

Figure 6.6: Residual manipulation results

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Frequency

Frequency

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 1 2 3 4 5 0 1 2 3 4 5
Time 4
x 10 Time x 10
4

(a) Spectrogram of a noise free speech signal. (b) Spectrogram of a noisy speech signal.

Figure 6.7: Spectrogram of speech, noisy speech and estimated speech signals

Group 844: 30th May 2005 31

 6.2. DSP ARCHITECTURE MAPPING

MIC

INPUT

Multi Channel AIC 23

LED
Serial Port Codec

OUTPUT

EDMA DSP
Controller Processor

M EMORY

Figure 6.8: Architecture mapping.

C version, the results are compared with Matlab results.

6.2 DSP Architecture Mapping

The basic architecture and functional units with functionalities of the DSP processor are
explained in appendix C. The different blocks on TMS320C6713 DSK board which are
used for LPC analysis implementation are shown figure 6.8. The processor and EDMA
(Enhanced Direct Memory Access) can directly access the memory. EDMA uses buffers
to read and process the data. The multi-channel serial port is used to communicate with
external peripherals like AIC23 codec and LED’s. The address and data buses are used to
communicate between the blocks [15].

6.2.1 ANSI-C on 6713DSK

The ANSI-C version, first executed on Matlab, then moved onto the 6713DSK with pos-
sible changes. The example programs were as for the reference to run on the 6713DSK.
To optimize the ’C’code on DSK kit, one should have good knowledge of memory allo-
cation, external peripheral port addresses and specific order of instruction execution. The
code is optimized with the help of example programs from code composer studio. The
Matlab code could not be worked out, so special Matlab functions were written like han-
ning window, framing, Levinson Durbin recusion algorithm, all pole filter, all zero filter.

Group 844: 30th May 2005 32

 6.2. DSP ARCHITECTURE MAPPING

The following is an example of all pole filter C code.

void AllPole Filter( float *In, /* (i) In[0] to In[lengthInOut-1] contain
filter input samples */
float *Coef,/* (i) filter coefficients (Coef[0] is assumed
to be 1.0) */
int lengthInOut,/* (i) number of input/output samples */
int orderCoef, /* (i) number of filter coefficients */
float *Out /* (i/o) on entrance Out[-orderCoef] to Out[-1]
contain the filter state, on exit Out[0]
to Out[lengthInOut-1] contain filtered
samples */
{
int n,k;
for(n=0;n<lengthInOut;n++)
*Out = Coef[0]*In[0];
for(k=1;k<=orderCoef;k++)
*Out += Coef[k]*In[-k];
}
Out++;
In++;
}
}
The aim is to implement the C programs on DSK. The TI C compiler(as with most com-
pilers) differentiate C-source labels from assembly source labels by prepending an ”_” to
all C labels as it generates assembly code. The many leader files were added to the main
program in C to run on DSK. CCS IDE (Integrated Development Environment) is made
up of two parts

• Edit (and Build):

Edit programs uses editor, to edit the programs and code generator tools to create
code.
• Debug (and Load):
Debug program is used to run the code.

Finally the program is loaded onto the DSK to test and confirm the operation of C pro-
gram.

Group 844: 30th May 2005 33

 6.2. DSP ARCHITECTURE MAPPING

Initialization
frame-counter=0
Px = Pmin

Load next frame

frame-counter+=1
Compute estimated
Speech spectrum
Compute power spectrum, Px
Sm = Xm *Q

Compute smoothed power

No spectrum, pxsmooth
Sm =Subf*Pn

if
Yes Compute Pmin
Sm ≤ Subf ∗ Pn

if
Compute Q factor Yes
Px ≤ Pmin

Osub=7-SNR*1/s
Pmin = Px
No
No
Osub=0
Pn =Omin*Pmin
if
SNR ≥ 35 Yes

if Yes
No Pn =0
Osub=7

if Yes SNR=35
SNR ≤ 0 No

Compute SNR

Figure 6.9: Flow chart of spectral subtraction.

Group 844: 30th May 2005 34

 6.2. DSP ARCHITECTURE MAPPING

Start

Load the speech signal (S)

Load the pink noise signal (n)

X=S+n

Circular buffer

Preemphasis filter

Sampling / Framing

Windowing (Hanning window)

Burst

Power spectrum
LPC Analysis filter (1) Spectral subtraction
Spectral manipulation

Residual signal (e) Autocorrelation of speech (ifft)

~
Residue manipulation ( e ) LPC (2)

Synthesis filter (Estimated speech)

Deemphasis filter

Save the noise reduced speech

End

Figure 6.10: Flow chart of Basic Algorithm

Group 844: 30th May 2005 35

 6.2. DSP ARCHITECTURE MAPPING

6.2.2 Optimizing the Code

The C program is loaded onto the DSK to confirm the functionality as mentioned in one-
day workshop student guide from TI. A new project directory is created on starting a new
project. A new .cdb (Configuration database) file was created to control the range of CCS
capabilities.
The created all files are added to project directory. On getting errors, break points were
used for the correction of the code until the execution was successful.
After execution, the results were stored in memory were compared with that of Matlab
results.
The Matlab converted C files and written C code along with Matlab code are placed in
CD-Rom.

Group 844: 30th May 2005 36

7.1 Results
Conclusions
7
The proposed algorithm was tested with noise free speech signals and additive and uncor-
related pink noise. After many experiments with different values, the smoothing factor of
the power spectrum was set to α=0.95. To smooth the minimum spectrum estimate, the
pole was set at γ = 0.9. The other parameters komin = 1.5, ksub f = 0.02, and D = 100
were set for the spectral subtraction. After spectral subtraction, the spectral manipulation,
i.e. the envelope estimation, smoothing in both time and frequency domains and resid-
ual manipulations were done with suitable filter. The estimated noise reduced speech
was observed with increased value of SNR at 0dB and 8dB. The perceptual quality was
enhanced and at the same time intelligibilty was maintained.
The second part of the project was about the study and the implementation onto TMS320C6713
DSK board. C code was written for LPC analysis and synthesis. Individual functions
were tested onto the DSK board. Working environment of DSK was determined and the
results observed were similar to that of the Matlab simulations results.

7.2 Conclusion

To conclude from the implementation results, the spectral subtraction method was used
for noise reduction in speech. It is used as a good noise power estimator. More analysis
was conducted using this technique and resulted in effective outcome. This technique
helped in increase the signal-to-noise ratio (SNR) and perceptual quality. Spectral ma-
nipulation was also performed, but not much analysis was done, as not much change was
observed in the result.
Residual manipulation played major role in order to maintain the intelligibility. This was
performed with the burst based execution in order to produce better reconstruction of
speech signal at synthesis filter.
It is recommended that new techniques are needed in order to suppress the noise in the
residual signal along with burst based execution and also in spectral manipulation.
The C-code was written in two methods. The first method: Matlab converted C code was
able to built on the board but while execution was terminated. The second method: The
written C code gave good performance compared to that of Matlab generated code. The
code was built and executed successfully.

Group 844: 30th May 2005 37

 Linear Predictive
Coding A
A.1 Introduction

Linear Predictive Coding (LPC) is a method used for estimating speech parameters like
pitch, formants, spectra, and vocal tract. LPC technique is based on linear predictive
analysis. LPC removes redundancy in speech signals by forming/exploiting correlation
in the speech. The analysis results in a set of AR coefficients, also known as whitening
filter coefficients. The speech signal is passed through the whitening filter to obtain the
residual signal, which looks like spectrally white. Finally the residual is encoded and
transmitted together with the AR coefficients. This chapter describes the basic overview
of linear prediction and different methods of approach to find the prediction coefficients
and error minimization.

A.2 Linear Prediction

The basic principle of linear prediction is to predict a future value of a stationary discrete-
time stochastic process, from a set of past or observed values [1, e.3,p.241]. The block
diagram of linear prediction is shown in figure A.1.

Prediction Residual
Linear x̂[n] -
x[n] z −1 predictor Σ e[n]
+

Figure A.1: Block diagram of Linear Prediction

As shown in the figure x[n] is a signal, considered as an input to a linear filter. The output
of the predictor is denoted by x̂[n]. It is the estimation of the signal x[n]. The error e[n]
is the difference between x̂[n] and x[n]. This is called linear estimation problem. In such
a problem the selection of a filter is designed to minimize the error e[n]. Now the aim
is to estimate future values from observed values of x[n], then the estimation called as
linear prediction. To estimate future values of x[n], the estimation is just obtaining the
linear prediction coefficients which are used in speech coding. A mathematical model of

Group 844: 30th May 2005 38

 A.3. ALL POLE MODEL

linear prediction is described in-order to obtain the predictor coefficients and minimize
the prediction error.
Consider a system with output x[n] with some unknown input u[n]. A linear prediction of
order p is to estimate the value x[n] as a linear combination of the P previous observations
x[n − 1], x[n − 2],.... x[n − p]. So the following equation is [6]:

p q
x[n] = − ∑ ak x[n − k] + G ∑ bl u[n − l], (A.1)
k=1 l=0

where ak , bl , and gain G are parameters of the system. Equation A.1 indicates that
output x[n] is a linear function of present, past inputs and past outputs. Then the signal
x[n] is predictable from linear combination of past outputs and inputs. The equation A.1
in frequency domain the transfer function is H(z) of the system.
X(z) B(z)
H(z) = =G , (A.2)
U(z) A(z)
where X(z) and U(z) are the z-transform of x[n] and u[n]. H(z) is the pole-zero model.
The pole-zero model is also called as autoregressive moving average (ARMA) model.
A system with pole-zero model the estimation of the parameters is inherently nonde-
terministic and nonlinear. But the all-pole model can approximate the speech spectrum
estimation closely and simply. So all-pole model is prefered in LPC [6].

A.3 All Pole Model

A system only with poles, that is bl = 0, 1 ≤ l ≤ q, and b0 = 1 in equation A.1, then the
system is referred to the all-pole model or auto-regressive (AR) model. Then the equation
A.1 reduce to:

p
x[n] = − ∑ ai x[n − i] + Gu[n], (A.3)
i=1

where G is the gain factor. The transfer function H(z) reduce to an all-pole transfer
function.
X(z) 1 1
H(z) = =G p =G . (A.4)
U(z) 1 + ∑i=1 ai z −1 A(z)
If the current input signal is unknown, which is the common case in most of the ap-
plications, then it can only approximately predict the signal x[n] from a linear weighted
summation of past samples. This approximation is denoted as x̂[n] and is expressed as:
p
x̂[n] = − ∑ ak x[n − k]. (A.5)
k=1

The error e[n] between the actual value x[n] and the predicted value x̂[n], is:
p
e[n] = x[n] − x̂[n] = x[n] + ∑ ak x[n − k], (A.6)
k=1

Group 844: 30th May 2005 39

 A.4. AUTOCORRELATION METHOD

where e[n] known as the residual. The prediction coefficients a k should be selected to in
order to minimize the total squared error. To obtain the coefficients the method of least
squares is applied. Then the equation is:
h p i2
En = ∑ e2n [n] = ∑ x[n] + ∑ ak x[n − k] , 0 ≤ n ≤ N − 1. (A.7)
n n k=1

The error is minimized by setting the error:

dEn
= 0, 1 ≤ i ≤ p. (A.8)
dai

From the equations A.7 and A.8 gives set of equations:

p
∑ ak ∑ x[n − k]x[n − i] = − ∑ x[n]x[n − i], 1 ≤ i ≤ p. (A.9)
k=1 n n

Equation A.9 is called least squares terminology. The equation A.9 forms a set of p linear
equations, that can be solved to find the predictor coefficients a k , 1 ≤ i ≤ p and also to
minimizes the En (equation A.7). The total minimum squared error denoted by E p , is
obtained by substituting the equation A.7 in A.9.
p
E p = ∑(x[n])2 + ∑ ak ∑ x[n]x[n − k]. (A.10)
n k=1 n

Now, the coefficients ak are obtained by solving equation A.10. Equation A.10 can be
solved using two different approaches. One is the autocorrelation and the another one is
covariance method.

A.4 Autocorrelation Method

In this method, the assumption is that the error in euqtion A.7 is minimized over the
infinite duration. The signal is considered a finite duration (N samples) to find the au-
tocorrelation [6]. The autocorrelations are calculated from the signal using available N
samples. Now, to solve the prediction coefficients a k , the equations A.9 and A.10 in-
terms of autocorrelation functions are:

p
∑ ak R(i − k) = −R(i). (A.11)
k=1
p
E p = R(0) + ∑ ak R(k), 1 ≤ i ≤ p. (A.12)
k=1

Group 844: 30th May 2005 40

 A.5. COVARIANCE METHOD

The equation system can be expressed with the matrix:

    
 R(0) R(1) R(2) . . . R(p − 1)   a1   R(0) 
    
 R(1) R(0) R(1) . . . R(p − 2)   a2   R(1) 
    
    
 R(2) . . . R(p − 3)   a3   R(2) 
 R(1) R(0)  =  (A.13)
 .. .. .. .. ..   ..  
 .. 
 .   
 . . . .  .   . 
    
R(p − 1) R(p − 2) R(p − 3) . . . R(0) ap R(p)

where R(i − k) is a symmetric autocorrelation matrix and R(i) is an autocorrelation func-

tion of signal x[n], it is also an even function. So this method is called the autocorrelation
method. Here signal x[n] is multiplied by a window, the signal x[n] = 0 outside the win-
dow (0 ≤ n ≤ N − 1) [11].The windowed signal spectrum is smoothed version of original
signal and the smoothness depends on the shape and size of the window. Because of win-
dowing the LPC coefficients results a smoothed spectrum. This method guarantees stable
filters [3]. The equation A.13 is solved by taking the advantage of the symmetry and the
fact that the elements across the diagonal are all identical (Toeplitz matrix), so the more
efficient Levinson-Durbin algorithm can be used to compute the matrix inversion.

A.5 Covariance Method

This is an alternative method to solve the predictor coefficients a k , 1 ≤ k ≤ p and also to

minimize the error En . This method use much smaller samples compared to autocorrela-
tion method, to minimize the error E n over a finite duration 1 ≤ n ≤ N − 1. The equations
A.9 and A.10 in terms of covariance functions are [6]:

p
∑ ak φki = −φi0 , 1≤i≤ p (A.14)
k=1
p
E p = φ00 + ∑ ak φ0k , (A.15)
k=1
where
N−1
φik = ∑ X[n − i]x[n − k], (A.16)
n=0
where φki is a symmetric covariance matrix. By solving the above covariance normal
equation set in matrix form as:
    
 φ1,1 φ1,2 φ1,3 . . . φ1,p   a1   φ1,0 
    
 φ2,1 φ2,2 φ2,3 . . . φ2,p   a2   φ2,0 
    
    
 φ . . . φ3,p   a3   φ3,0 
 3,1 φ3,2 φ3,3  =  (A.17)
 . .. .. .. ..  
 ..  
 .. 
 . .   
 . . . .  .   . 
    
φ p,1 φ p,2 φ p,3 . . . φ p,p ap φ p,0

Group 844: 30th May 2005 41

 A.6. LEVINSON DURBIN ALGORITHM

The matrix equation A.17 is symmetric but not Toeplitz matrix (the elements across the
diagonal are different). This method cannot guarantee stable filters. There is no efficient
algorithm to make the matrix inversion for covariance matrix equation.

A.6 Levinson Durbin Algorithm

The Levinson Durbin (L-D) algorithm describes a direct method for computing the pre-
dictor coefficients ak and mean square error En for a order p by solving the augmented
Wiener-Hopf equations. The method is recursive in nature and by using Toeplitz struc-
ture of the correlation matrix (equation A.13) of a tap input of the filter. This known as
Levinson Durbin algorithm [1, e.3,p.198].
The Toeplitz matrix equation A.13 is solved by using L-D algorithm, which is efficient
method to exist to invert such matrices. To determine the values of the linear prediction
parameters at iteration i by aik and the error Eni for 1 ≤ i ≤ p. The recursive procedure is:
Initially E0 = R(0) and a0 = 0.

(i−1)
R(i) + ∑i−1
j=1 a j R(r − j)
ki = . (A.18)
E i−1
The parameters ki are known as the reflection coefficients.

aii = ki . (A.19)

(i−1) (i−1)
aij = a j + ki ai− j , 1 ≤ j ≤ i−1 (A.20)

Ei = (1 − Ki2 )Ei−1 . (A.21)

The equations A.19 and A.21 are solved recursively in each iteration. In each cycle the
iteration i, the coefficients aik for 1 ≤ i ≤ p gives the optimal pth order linear predictor and
error En is reduced by a factor of (1 − ki2 ) [6]. The filter with L-D algorithm produces is
in minimum phase and stable. L-D algorithm is computationally efficient.

Group 844: 30th May 2005 42

 Speech and Noie
B
This chapter is introduced to describes the properties of speech and noise. In the speech
section it describes about formants and pitch period. In the noise section, different noise
types and its charectaristics are described.

B.1 Speech

Speech is used to communicate information from a speaker to a listen. The human speech
production begins with an idea or thought that the speaker wants to convey to a listener.
The speaker produces an acoustic sound pressure wave by a series of neurological process
and muscular movements that is received by a listener auditory system.
Speech signals are non-stationary and at best they can be considered as quasi-stationary,
over the short segments, typically 5-30 msec. The statistical and the spectral properties of
speech are thus defined over a short segment. Speech can be generally classified as voiced,
unvoiced or mixed. Voiced speech is quasi-periodic in the time domain and harmonically
structured in frequency domain while unvoiced speech is random like and broadband. In
addition, the energy of voiced segments is generally higher than the energy of unvoiced
segments.
The short time spectrum of voiced speech is characterized by its fine and formant struc-
ture. The fine harmonic structure is a consequence of quasi-periodicity of speech and
may be due to the vibrating vocal cords. The frequency of periodic pulses is referred to
as the fundamental frequency or pitch. The formant structure (Spectral envelope) is due
to the interaction of the source and vocal tract. The spectral envelope is characterized by
a set of peaks which are called formants. The formants are the resonant modes of vocal
tract. These formants are quite important both in speech synthesis and perception.

B.2 Noise

The noise can be defined as complex sound waves that are aperiodic, that is, the sound
waves with irregular vibrations and no definite pitch. In other words, noise is defined as
a unwanted signal that interferes with the detection of or quality of another signal [16].
The noise is classified into different colors according to their spectral properties. White
noise power density is constant over a finite frequency range. The next most commonly
used color is pink noise. Its frequency spectrum is not flat, but has equal power in bands
that are proportionally wide. Pink noise is perceptually white. That is, the human auditory

Group 844: 30th May 2005 43

 B.2. NOISE

system perceives approximately equal magnitude on all frequencies. The power density
decreases by -3 dB per octave (density proportional to 1/ f ). Brown noise is similar
to pink noise, but with a power density decrease of -6 dB per octave with increasing
frequency (density proportional to 1/ f 2 ) over a frequency range which does not include
DC. There are also many "less official" colors of noise such as red, orange, green and
black.

Group 844: 30th May 2005 44

 C DSP
This chapter includes DSP processor architecture, development tools and issues related
to real-time processing considered in this project. The TM320C6713 DSP starter kit with
high precision applications based on TI’s TM320C6000 floating point DSP generation.
The TMS320 DSP family offers the most extensive selection of DSPs available, with a
balance of general-purpose and application-specific processors to suit user needs. There
are distinct Instruction Set Architectures that are completely code-compatible within plat-
forms.

C.1 TMS320C6000 DSP platform

The C6000 DSP platform offers fast DSPs running at clock speeds up to 1 GHz. The
platform consists of the TMS320C64x and TMS320C62x fixed-point generations as well
as the TMS320C67x floating-point generation. The C6000 DSP platform’s performance
ranges from 1200 to 8000 MIPS (Million Instructions per Second) for fixed-point and 600
to 1350 MFLOPS (Mega Floating Point Instructions per Second) for floating point [15].
Basic C6000 CPU Architecture:

• Functional Units
• Register File
• Memory and Peripheral

C.2 Functional Units

It contains eight independent functional units as shown in the figure C.1. All eight func-
tional units can receive their own 32 bit instruction on every cycle, i.e. it can execute
eight instructions in parallel.

• .D unit (.D1,.D2) A 32-bit loads and stores, add, subtract, linear and circular ad-
dress calculations.
• .M unit (.M1,.M2) It performs 16x16-bit integer or 32x32-bit floating point multi-
ply operations in the hardware.
• .L unit (.L1,.L2) A 32/40-bit arithematic and compare, 32 bit logic operations. It
performs converstion operations.

Group 844: 30th May 2005 45

 C.3. REGISTER FILE

Memory

A0 B0

A1 .D1 .D2 B1

.M1 .M2

....
....

.L1 .L2

A14 .S1 .S2 B14

A15 B15

Controller / Decoder

Figure C.1: Architecture mapping.

• .S unit (.S1,.S2) A 32/40-bit shifts and 32-bit bit fields, branches, constant gen-
eration compare, reciprocal and reciprocal square-root, absolute value operations,
conversion. Register transfers to/from control register file.

C.3 Register File

The variables operated upon by the CPU are stored in register file. There are two reg-
ister files. Register file "A" (A0-A15/31) and Register file "B" (B0-B15/31) of 16 or 32
registers each, depending upon which C6000 CPU is using.

C.4 TMS320C6713 DSP

The TMS320C6713 is the floating-point DSP generation in the TMS320C6000 DSP plat-
form. The C6713 DSP also features a two-level cache and VLIW (very-long instruction
word) architecture. The C6713 DSP operating at 225 MHz, delivers up to 1350 mil-
lion floating-point operations per second (MFLOPS), 1800 million instructions per sec-
ond (MIPS), and with dual fixed-/floating-point multipliers up to 450 million multiply-
accumulate operations per second (MMACS). The C6713 DSP has sufficient bandwidth
to support all 16 serial data pins transmitting a 192 kHz stereo signal.

Group 844: 30th May 2005 46

 C.5. C6713 MEMORY AND PERIPHERALS

Address TMS320C6713 C6713 DSK

0x00000000 256 kB Internal

Program / Data

0x00030000 8MB SDRAM

Reserved Space
Or
Peripheral Regs 0x807FFFFF
256 kB Flash
0x80000000 128MB External Memory
(EMIF CEO) 0x90080000
CPLD
0x90000000 128MB External Memory
(EMIF CE1)

0xA0000000 128MB External Memory Available via

(EMIF CE2) Daughter card

connector
0xB0000000 128MB External Memory
(EMIF CE3)

Figure C.2: C6713 Memory map.

C.5 C6713 Memory and Peripherals

The c6x family of DSPs has a single large 32-bit address space. The address space is
split between on-chip memory, on-chip peripheral registers and external memory. All
memory is byte addressable and program code and data can be mixed freely. The C6713
also has 4kB program and data caches to improve performance when accessing external
code and data [15]. The figure C.2 memory map for the C6713 DSK showing how the
address space is used. The internal memory starts at the begining of the address space
with most of the either reserved or used for peripheral registers. The EMFI starts at
address 0x80000000 and spans the next 1 GB of the address space. It is divided into 4
equally sized regions each with a dedicated chip enable signal (CE0-CE3). The on-board
memory, programmable CPLD (complex programmable logic device) registers and add-
on daughter cards are all connected through EMIF (external memory interfaces).

C.6 TMS320C6713 DSP Starter Kit (DSK)

The TMS320C6713 DSP Starter Kit (DSK, developed for high precision applications
based on TMS320C6000 floating point DSP generators. Like audio, medical imaging,
test and instrumentation. The C6713 DSK includes 8MB of on-board SDRAM and an
emulation header and 12C interfaces.
The DSK includes the Fast Run Time Support libraries and utilities such as Flash burn
to program flash, Update Advisor to download tools, utilities and software and a power
on self test and diagnostic utility to ensure the DSK is operating correctly. The hardware

Group 844: 30th May 2005 47

 C.7. TOOLS AND SOFTWARE

features of the TMS320C6713 DSK board include:

• C6713 DSP Development Board with 512K Flash and 8MB SDRAM
• High-quality 24-bit stereo codec
• Four 3.5mm audio jacks for microphone, line in, speaker and line out
• Expansion port connector for plug-in modules

C.7 Tools and Software

For DSP product development, the TMS320 DSP family is supported by user eXpress-
DSP Real-Time Software Technology that includes Code Composer Studio integrated de-
velopment environment, DSP/BIOS Real-time software kernel, TMS320 DSP Algorithm
Standard.

C.8 Algorithm Standard

The TMS320 DSP Algorithm Standard is a single, standard set of coding conventions
and application programming interfaces (APIs) for algorithm creators. The standard in-
cludes algorithm programming rules that enable interoperability between different types
of algorithms.

C.9 Terminology

Below are some of the more common terms that encounter when discussing the TMS320
DSP Algorithm Standard.

• Algorithm: A module of code that consumes a data stream, processes it, and out-
puts a resultant stream. Examples include vocoders, modems, audio compression,
video decompression, etc.
• Reference Framework : The "glue" code that holds together the drivers, the al-
gorithms, resource managers, and DSP kernel. Reference Frameworks start out
as application-agnostic. Upon the addition of application-specific algorithms, the
Framework takes on an application-specific nature.
• DSP Kernel: A low-level software layer that provides hardware abstraction and
manages low level physical resources. It provides threading; interrupt support,
pipes, signals, and several other functions. In addition, DSP/BIOS (Basic Input
Output System) offer data logging and statistical accumulation that enable real-
time analysis of the system.

Group 844: 30th May 2005 48

 C.9. TERMINOLOGY

• Application: It depends on the use of some or all of the other components. If a user
writes all the code from scratch including a kernel, algorithms, and a framework,
then the entire software system may be described as the application. However, in
an environment where DSP/BIOS, a reference framework, and COTS (Commercial
off-the-shelf) algorithms have been deployed, the application programmer uses the
APIs (Application Program Interface) for the controlling framework.

Group 844: 30th May 2005 49

 Bibliography
[1] Simon Haykin:, "Adaptive filter theory", Edition.4,Prentice-Hall, Inc, 2002,
ISBN:0-13-048434.

[2] Ljung:, "System identification. Theory for user", Edition.2, Prentice Hall, 1999,
ISBN:0-13-881640-9.

[3] John R. Deller, Jr. and John H.L Hansen and John Proakis:, Edition.2, "Discrete-
Time Processing of Speech Signals", Wiley-Interscience-IEEE, 0-7803-5386-2.

[4] A.M.Kondoz:, Digital Speech, Coding for low bit rate communication sys-
tems,Edition.2, Wiley and Sons, 1999, ISBN: 0471623717.

[5] Mohammad.M.A.khan:, "Coding of excited signal in a waveform interpolations",

McGrill University, Canada, july,2001.

[6] John Makhoul:, Linear Prediction: A Tutorial Review, IEEE Transactions on Signal
Processing, 1975, Vol.SIG-63, No. 4, April

[7] J.D.Markel and A.H.Gray, K.S.Fu, W.D.Keidel, H.Wolter:, Linear Prediction of

Speech, Springer-Verlag, 1976, ISBN: 3540075631

[8] Martin, R.: "Spectral Subtraction Based on minimum Statistics", EUSIPCO-94, Ed-
inburgh, Scotland, 13.-16 September 1994, pp. 1182-1185

[9] Steven F. Boll: "Suppression of Acoustic Noise in Speech Using Specral Subtrac-
tion", Edition.2, IEEE Transactions on Acoustics, Speech, and Signal Processing,
volume.27, pages.113-120, april,1979, Isbn:0096-3518.

[10] A.Th.Schwarzbacher and J.Timoney:, "VLSI Implementation of Noise Canceller",

3rd Int. Symposium on communication Systems, July,2002.

[11] O’Shaughnessy:, "Linear predictive coding", IEEE potentials, volume.7, pages.29-

32, october-1998, ISBN:0278-6648.

[12] M. Berouti and R. Schwartz and J. Makhoul:, "Enhancement of Speech Corrupted

By Acoustic Noise", Acoustics, Speech, and Signal Processing, IEEE International
Conference on ICASSP ’79, vol.4, pages.208-211, april,1979.

Group 844: 30th May 2005 50

 BIBLIOGRAPHY

[13] Axel Jantsch and Shashi Kumar and Ahmed Hermani:, "The Rugby Model: A Con-
ceptual Frame for the Study of Modelling, Analysis and Synthesis Concepts of Elec-
tronic Systems", Design, Automation and Test in Europe Conference and Exhibition
1999. Proceedings, pages.256-262, 1999.

[14] http://burks.brighton.ac.uk/burks/foldoc/59/51.html, may,5,2005

[15] http://focus.ti.com/docs/prod/folders/print/tms320c6713.html, may,15,2005

[16] http://www.asha.org/public/hearing/disorders/noise.html, may,15,2005.

Group 844: 30th May 2005 51

.