Speaker Recognition System

Heli Patel (19BEC037), Hitarth Patel (19BEC039)

Department of Electronics and Communication

Institute of Technology ,Nirma University

Ahmedabad ,India
19bec037@[Link] 19bec039@[Link]

Abstract— Speaker Recognition system is important for

biometrics and recognizing the voice .Speech signal is a III. PRICIPLES OF SPEAKER RECOGNITION
complex signal as it has many transformations. It can
also be easily affected by various factors like The basic principles of speech recognition are:
background noise. Thus we need a device that overcomes
all such difficulties and recognizes the speech correctly. A. Speech Editing

In this process from the speech signal a set is recorded in

I. INTRODUCTION wave format and then on this set the editing is done. The
vector is divided into two vectors of equal length but
Identity verification is an important aspect in many fields opposite orders and with the help of MATLAB the file is
and speaker recognition also plays a part in it. It is important read and also played again in reverse order.
in privacy matters and also issues of identity theft and thus
helps in security. It can also be used to get access to homes,
accounts or other applications. Speaker recognition as the
name itself suggests is a method in which we recognize the B. Speech Degradation and Enhancement
spoken language. It is also very common in communications
through mobiles, telephones and other devices. It is also an
important development for blind and deaf people to easily To enhance or degrade any speech, noise plays a vital role
communicate. and thus it becomes essential to estimate noise. When there
is more noise the signal gets distorted and thus to remove it
II. VERIFICATION OF SPEAKER and prevent speech degradation, Gaussian noise is added to
the signal. For speech enhancement we remove this
Gaussian noise by first applying fft on the signal to convert
Any speech recognition has 5 basic steps to be followed: it into frequency domain. Now those components which
have higher frequencies are removed by butterworth low
1. Digital Speech Data Acquisition pass filter (as order n is increased, behaves as an ideal low
2. Feature Extraction pass filter and has better capability to filter Gaussian noise).
3. Pattern Matching The signal is then scaled and plotted.
4. Accepting/Rejecting Decision
5. Enrolment to generate speaker reference models.
C. Pitch Analysis
When digital speech is received each interval of speech is
mapped on a multidimensional feature space and is then Pitch analysis is the ordering of sounds on frequency related
compared with speaker models by pattern matching. For scale. It identifies state of speech i.e. whether the speech is
each vector or its sequence the pattern matching gives us a happy, sad, neutral, or any other emotions. This can be done
match score. This match score is what compares the by calculating the average and then noticing the differences.
similarity of the received data to that stored in the speaker
and finally the decision of accepting or rejecting is made.
D. Format Analysis
The different pattern matching models that are used in
speech recognition are:
Format analysis can be done by calculating the difference
1. Dynamic Time Warping (DTW) between the vector peak positions, and the vector position of
2. Hidden Markov Model (HMM) the peaks in power spectral density. The speaker recognition
systems should have enrolmentsessions or training phase
3. Artificial Neutral Networks
and operation sessions or testing phase. In the 1 st phase the
4. Vector Quantization (VQ) speaker sends a sample of its speech to train the reference
model while in testing phase, the input signal is matched
with the stored reference model. And then recognition
decision is made.

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

 IV. SPEECH PROCESSING characteristics which include: speaking rate, prosodic effects
and dialect.
This is the method of extracting the desired data from the
speech signal and in order to process a speech signal the top C. LP
most priority is that the signal must be in digital form.
Keep your text and graphic files separate until after the LP model is an all-pole model in which the signal sn is a
text has been formatted and styled. Do not use hard tabs, and
linear combination of its past values and a scaled present
limit use of hard returns to only one return at the end of a
paragraph. Do not add any kind of pagination anywhere in input. Mathematically it can be represented as:
the paper. Do not number text heads-the template will do that 𝑝
for you. sn = -∑𝑘=1 𝑎 𝑘 .𝑠𝑛−𝑘 + 𝐺. 𝑢𝑛

Here, sn is the current output, p the prediction order, a k the

predictor coefficients, sn-k the past outputs, G the gain
scaling factor and u n the present input.

If we approximate LP depending only on the past samples:

𝑝

𝑠̂𝑛 = ∑ 𝑎 𝑘 . 𝑠𝑛−𝑘
𝑘=1

Now if we define e n as the prediction error then we get:

𝑝

𝑒𝑛 = 𝑠𝑛 − 𝑠̂𝑛 = 𝑠𝑛 + ∑ 𝑎 𝑘 . 𝑠𝑛−𝑘
𝑘=1

Consider MSE to be E,
𝑝
A. Speech Signal Acquisition
𝐸 = ∑ 𝑒𝑛2 = ∑ [𝑠𝑛 + ∑ 𝑎 𝑘 . 𝑠𝑛−𝑘 ] 2
𝑛 𝑛
For the digital transformation of the speech signal first of all 𝑘=1
the speech signal i.e. the acoustic pressure waves are
The minimum MSE
collected by a microphone or telephone and converted into
analog signal. The then converted signal is passed through
an antialiasing filter which limits its bandwidth to ∑𝑝𝑘=1 𝑎 𝑘 . ∑𝑛 𝑠𝑛−𝑘𝑠𝑛−𝑖 = − ∑𝑛 𝑠𝑛 𝑠𝑛−𝑖
approximately the Nyquist Rate. Now an A/D converter is
used to finally convert the signal into digital. where i= 1,2,….p
.
B. Speech Production This results in the autocorrelation method of LP. Its time
averaged estimates at lag τ are:
The vocal tract helps produce speech. There are different
types of excitation that occurs upon the different ways of Rτ = ∑𝑁−1−𝜏
𝑖=0 𝑠(𝑖) . 𝑠(𝑖 + 𝜏)
airflow in the vocal folds. Some of these are Phonated
excitation, Whispered excitation, Frication excitation, This method gives:
Compression excitation,etc. All these different excitations
have different speech models for the process of speech 𝑅0 𝑅1 𝑅2 ……….. 𝑅𝑝−1
𝑎1 𝑅1
recognition. 𝑅1 𝑅0 𝑅1 𝑅𝑝−2
……….. 𝑎2 𝑅2
Mostly the system used for voice recognition use features 𝑅2 𝑅1 𝑅0 𝑅𝑝−3
that are developed from the vocal tract. And thus in voice
……….. 𝑎3 𝑅3
. . =− .
recognition, the shape of the vocal tract also plays a major .
role. This shape can be estimated from the spectral shape of . .
. . .
the voice signal. The excitation as discussed above is a .
speaker-dependent information and it is not the only form of [𝑎 𝑝] 𝑅
[ 𝑝]
[ 𝑅𝑝−1 𝑅𝑝−2 𝑅𝑝−3 ……….. 𝑅0 ]
speaker-dependent information available. Vital capacity,
maximum phonation time, phonation quotient and glottal air
flow are also under this category. This system of equations are solved by Durbin’s recursive
These above characteristics are those which are physical but algorithm.
there are also some of the characteristics that help
E0 = R 0
distinguish between speakers and those are the learned
 (𝑖−1)
𝑘 𝑖 = −[𝑅𝑖 + ∑𝑖−1
𝑗=1 𝑎𝑗 𝑅𝑖−𝑗 ]/𝐸𝑖−1 LAR can be defined as log of the ratio of adjacent cross
( 𝑖) sectional areas of the cylinders:
𝑎 𝑖 = 𝑘𝑖
(𝑖) ( 𝑖−1) ( 𝑖−1)
v i=1,2,……p
𝑎𝑗 = 𝑎𝑗 + 𝑘 𝑖 𝑎𝑖−𝑗 𝑣1 ≤ 𝑗 ≤ 𝑖 −1 𝐴𝑖+1 1 + 𝑘𝑖
𝑔𝑖 = log[ ] = log[ ] = 2 tanh−1 𝑘 𝑖
𝐸𝑖 = (1 − 𝑘 2𝑖 )𝐸𝑖−1 } 𝐴𝑖 1 − 𝑘𝑖
(𝑝)
𝑎𝑗 = 𝑎𝑗 𝑣1 ≤ 𝑗 ≤ 𝑝
3. Arcsin Reflection Coefficients
This shows that any signal can be represented as linear
predictor and the corresponding LP error. Now, Here k i =1 so that singularity of LAR is prevented
𝑝 𝑔𝑖′ = sin−1 𝑘 𝑖
sn = -∑𝑘=1 𝑎 𝑘 .𝑠𝑛−𝑘 + 𝑒𝑛

Transfer function of LP is: 4. LSP Frequencies

𝑆(𝑧) 𝑍[𝑠𝑛] These are the p zeroes of A(z) mapped on a unit circle in z
𝐻 (𝑧) ≡ ≡
𝑈(𝑧) 𝑍[𝑢𝑛] plane by the help of pair of auxiliaries(p+1) order
polynomials.
𝐺 𝐺
𝐻 (𝑧) = 𝑝 ≡
1 + ∑𝑘=1 𝑎 𝑘 𝑧−𝑘 𝐴(𝑧) A(z) = ½[A(z)+z-(p+1)A(z-1)+ A(z)-z^-(p+1)A(z-1)]
A stable LP filter has all its poles inside the unit circle and
A(z)is the p th order inverse filter or whitening filter.
If the voice signal fits the model then an impulse train thus the inverse filter has no poles or zeroes outside the
circle thereby giving minimum phase inverse. LSP
repeating at the rate of vocal fold vibration is the residual
and so it faces many errors. In time domain most of the representation of LP has direct frequency-domain
energy lost takes place near these pitch peaks. interpretation.

Some of the features that help speech coding and

recognition are:

1. Reflection Coefficients

These are the k i variables and can be obtained by backward

recursion of LP coefficients.
(𝑝)
𝑎𝑗 = 𝑎𝑗

( 𝑖) Frequency Response
𝑘𝑖 = 𝑎𝑖
(𝑖) (𝑖) (𝑖)
( 𝑝) 𝑎𝑗 + 𝑎 𝑖 . 𝑎 𝑖−𝑗 D. Mel-Warped Cepstrum
𝑎𝑗 = 𝑣1 ≤ 𝑗 ≤ 𝑖 −1
1 − 𝑘 2𝑖
i= p,p-1,……1 This feature does not require LP analysis. Here the signal is
windowed and then fft is taken. After that magnitude is
2. Log Area Ratios calculated and its log is done. The frequencies are warped
according to mel scale and at the end inverse fft is done.
Here the vocal tract is considered as a series of It is beneficial as it has well linear combination of Gaussian
cylindrical acoustic tubes. At each junction there is a densities and it gives better performance.
possibility of impedance mismatch or analogous
difference. The wave transmitted at each boundary also
reflects back some of its portion and these can be termed
as k i. If we consider that the tubes have equal length then
the time taken by the sound to travel through all will be V. FEATURE SELECTION AND MEASURES
the same. So this allows simple z transformation for
digital filter simulation. To derive the reflection It includes:
coefficients:
A0 = 0
Ap+1>>Ap A. Traditional Feature Selection

𝐴𝑖+1− 𝐴𝑖 Here from a set of variables, feature vector vector is

𝑘𝑖 = v i=1,2,….p extracted and the conversion of them into feature vectors is
𝐴𝑖+1+ 𝐴𝑖
feature selection. The main aim is to find a transformation of
the vectors to a relatively low dimensional feature space. To The population entropy H for an ensemble of pattern vectors
reduce the dimensionality, principal component analysis and having pdf p(x) is:
factor analysis are done. By linear transformations the H= -E[lnp(x)]
feature can be divided by hyperplane . 1
=-∫𝑥 𝑝(𝑥) 𝑙𝑛𝑝(𝑥) 𝑑𝑥
If we have x variable with mean u x and covariance C x and m
x n transformation matrix A, p(x) ~ N(u x, Cx), y=Ax and For ith class entropy will be
p(y)~N(A u x,A CxA^T).
y=Ax 1
µ y = E[y]= E[Ax] =AE[x] = A µ x Hi = ∫𝑥 𝑝𝑖 (𝑥) 𝑙𝑛𝑝𝑖 (𝑥)𝑑𝑥

Cy = E[(y - µ y)(y- µ y)^T]=E[A(x - µ x)(A(x- µ x))^T]=AE[(x - Directed divergence is:

µ x)(x- µ x)^TA^T]=AC xA^T
1 𝑝𝑖( 𝑥)
I(i,j)= ∫𝑥 𝑝𝑖 (𝑥) ln ( ) 𝑑𝑥
It has normal density. Linear transformation is not always 𝑝𝑗( 𝑥)
useful and thus non-linear transformation comes in view. To
identify good features analysis of variance technique The divergence is:
(ANOVA) is used wherein in Fisher’s ratio (F ratio) is 1 𝑝𝑖( 𝑥)
Jij = ∫𝑥 [𝑝𝑖 (𝑥) − 𝑝𝑗 (𝑥) ]ln ( ) 𝑑𝑥
calculated. F = variance of speaker means/average 𝑝𝑗( 𝑥)
intraspeaker variance. However, to be helpful F ratios of
many different combinations are to be measured will
becomes a challenge.
VI. PATTERN MATCHING
B. Mean and Covariance Estimation
Pattern Matching involves computation of a match score,
which is referred to as similarity of input vector to some
The covariance of a sample: model. Voice model of the user is generated and stored on
the recognition system. Algorithm compares incoming
𝑁
1 signal from the speaker with the model of the claimed user.
𝐶̂ = ∑(𝑥𝑖 − µ)(𝑥𝑖 − µ) 𝑇 Models are of two types:
𝑁−1
𝑖=1 1. Stochastic models
2. Template models.
The sample mean:
In case of stochastic models, matching of pattern is
𝑁+1 probabilistic and it results in a measure of likelihood of the
1 1
µ̂
𝑁+1 = ∑ 𝑥𝑘 = µ̂𝑁 + (𝑥 − µ̂𝑁 ) observation given the model. In case of template models,
𝑁+1 𝑁 + 1 𝑁+1 matching of the pattern is deterministic. Observation is
𝑘=1
assumed as an imperfect duplication of the template. The
Sample covariance matrix recursion alignment of observed frames to template frames is chosen
𝑁+1
1 to reduce a distance [Link] can be
̂
𝐶𝑁+1 = ∑ (𝑥𝑘 − µ̂ 𝑁+1 )(𝑥𝑘 − µ̂𝑁+1)
𝑇
approximated in template based models by :
𝑁
𝑘 =1
𝑁−1 1 𝐿 = 𝑒 −𝑎𝑑
= ̂𝑁 +
𝐶 . (𝑥
𝑁 𝑁 +𝑇1 𝑁+1 Here a is a positive constant.
− µ̂𝑁 )(𝑥𝑁+1 − µ̂𝑁 )

This shows the speaker identity. A. Template Models

It consists of single template 𝑥̅ , a model for a frame of

C. Divergence Measure speech. Match score between claimed user and input signal
𝑥𝑖 from another unknown user is indicated by 𝑑(𝑥𝑖 , 𝑥̅ )
It is the measure of distance / dissimilarity between 2 classes Distance measures between 𝑥𝑖 and 𝑥̅ is given by,
based upon information theory. According to the Tou and
Gonzalez’s derivation: 𝑑 (𝑥𝑖 , 𝑥̅ ) = ( 𝑥𝑖 − 𝑥̅ ) 𝑇 𝑊(𝑥𝑖 − 𝑥̅ )
p i(x)=p(x|ω i)
p j(x)=p(x|ω j) Here 𝑊 is a weighting matrix. Distance is Euclidean if 𝑊 is
identity matrix. Distance is Mahalanobis distance if 𝑊 is the
inverse covariance matrix corresponding to 𝑥̅ . Model for
The likelihood ratio is given by:
claimed speaker can be the mean of a set of training vectors
uij= ln(pi(x)/ pj(x))
 𝑁
1
𝑥̅ = µ ∑ 𝑥𝑖
𝑁
𝑖=1

VII. CLASSIFICATION AND DECISION

THEORY
Decision to accept or reject speaker or request again for
utterance is made after verification of score between the
input speech vector and claimed speaker’s voice vector.
Accepting or Rejecting decision process may be continue,
reject or accept, time out. This procedure is sequential
hypothesis problem. H y p o thetical R o c

A. Hypothesis Testing
CONCLUSION
Hypothesis testing involves choosing between the two
hypothesis : Speaker recognition is the utilization of a machine to
1. User is claimed speaker perceive an individual from a verbally expressed expression.
2. User is not the claimed speaker. Speaker recognition frameworks can be utilized either to
recognize a specific individual or to confirm a people
Consider that the hypothesis of user who is not the claim ed guaranteed personality. The fundamentals of speaker
speaker is H0 and that of a user who is claimed speaker is recognition and measures for speaker recognition were
H1. Consider conditional density function generated by the introduced and contrasted and customary ones utilizing
user who is not the claimed speaker is p(z| H0) and by the speaker-segregation standards. Speaker recognition systems
user who is claimed speaker is p(z| H1) having the match can be designed by matching the pattern with hidden
score of [Link] that true conditional score densities Markov model vector quantization , MFCC etc.
are known for both user who is not the claimed speaker as
well as for the user who is the claimed speaker then, Bayes’s
test is based up on likelihood ratio for the speaker, λ(z)
MATLAB CODES
𝑝(z| 𝐻0 )
λ(z) = • Train code
𝑝(z| 𝐻1)
fs=8000;
If the overlap between pdf’s of two score is small, then
% Sampling rate
probability of error is also small. Overlap between two pdf’s
nbits=16;
is given by,µ 0
nChannels=1;
(µ0 − µ1 )2 duration=5;
𝐹=
𝜎2 % Recording duration
Where µ0 and µ1 are means and 𝜎 is a variance. arObj=audiorecorder(fs, nbits, nChannels);
Likelihood ratio can be determined by choosing threshold fprintf('Please Press any Key on the Keyboard, to
X, START RECORDING:___', duration); pause
≥ 𝑋, 𝑐ℎ𝑜𝑜𝑠𝑒 𝐻0
if λ(z) { fprintf('Please wait while it is Recording');
< 𝑋, 𝑐ℎ𝑜𝑜𝑠𝑒 𝐻1 recordblocking(arObj, duration);
fprintf('Your Voice has now been RECORDED\n');
B. ROC fprintf('Please Press any Key on the Keyboard, to
PLAY RECORDING'); pause;
fprintf('\n');
One error type can be reduced by increasing the other type play(arObj);
of error. Relation between false acceptance and false fprintf('Plotting the waveform\n');
rejection is a function of decision threshold. For any of the y=getaudiodata(arObj);
system, ROC can be traversed by making changes in the % Fetching the audio sample data
threshold of acceptance likelihood ratio. Straight line ROC
plot(y);
indicates that product of probability of False Acceptance % Plotting the waveform
and probability of False Rejection is constant and equal to figure;
Equal Error Rate which is the value at which the value of
f = Simplefft(y);
False Acceptance and False Rejection are equal. % Autocorrelation
ms2 = fs/500;
ms20 = fs/50; f = Simplefft(y);
r = xcorr(y, ms20); load database
d = (-ms20:ms20)/fs; D=[];
plot(d, r); for(i=1:size(F,1))
title('Autocorrelation Form'); d=sum(abs(F(i)-f));
xlabel('Delay (s)'); D=[D,d];
ylabel('Correlation Coefficients'); end
r = r(ms20 + 1 : 2*ms20+1); sm =inf;
[rmax, tx] = max(r(ms2:ms20)); ind = -1;
Fx = fs/(ms2+tx-1); for(i=1:length(D))
Fth= 180; if(D(i)<sm)
%% Threshold frequency is 180 Hz sm=D(i);
if Fx> Fth ind=i;
disp('Speaker is Female!') end
gender = 'female'; end
else Identity_Number= C(ind);
disp(' Speaker is Male !') % disp("The Identity Number is;");
gender = 'male' Identity_Number
end
%% Saving the user data in the MATLAB Database • FFT
idno = input("Enter Corresponding Identity Number:");
try % DSP Voice Database Allowance via Correlation
load database %FFT Function
F=[F;f]; function [xPitch]= Simplefft(y)
C=[C;idno]; F = fft(y(:,1));
G=[G;gender] %Plotting the function
save database22 plot(real(F));
catch m = max(real(F));
F = f; xPitch= find(real(F)==m,1);
C = idno;
G = gender;
save database F C G;
end
msgbox('Thank You, your Voice has been Registered')

• Test Code
fs=8000;
% Sampling rate
nbits=16;
nChannels=1;
duration=5;
% Recording the Voice Input duration
arObj=audiorecorder(fs, nbits, nChannels);
fprintf('Please Press any Key on the Keyboard, to
START %g Seconds RECORDING__', duration);
pause
fprintf('Please wait while it is Recording\n');
recordblocking(arObj, duration);
fprintf('Your Voice has now been RECORDED\n');
fprintf('Please Press any Key on the Keyboard, to
PLAY RECORDING');
pause;
fprintf('\n');
play(arObj);
fprintf('Plotting the waveform\n');
y=getaudiodata(arObj);
% Getting the Audio Sample Data
plot(y);
% Plotting the Waveform
figure;
% Extraction
 References
[1] Speaker Recognition: A Tutorial - Proceedings of the
IEEE ([Link])
[2] [Link]
ition/