Efficient Speaker Identification System using MFCC [600520]

Efficient Speaker Identification System using MFCC
and Similarity Measurements

A. MAAZOUZI1,*, N. AQILI1, A. AAMOUD1, M. RAJI1, A. HAMMOUCH1
1LRGE Laboratory, ENSET
Mohammed V University in Rabat
Rabat, Morocco
*Correspondance : [anonimizat]

Abstra ct— Among various techniques in user identification,
speaker identification concerns with finding the identity of a
person via voice. Most of speaker identification systems are based
on the computation of distance or likelihood. The identification
process depends on the number of feature vectors, their
dimensionality and the number of speakers. This research aims
to develop a system able to identify a person from a sample of his
speech. Recognition relies on a text -dependent system using
English words as a password. Speech features are extracted using
Mel Frequency Cepstral Coefficients (MFCCs). The recognition
is based on Discrete to Continuous algorithm. Experimental
result s demonstrated that the proposed system gives a good
accuracy rate.
Keywords — Speake r identification ; verification ; text-
dependent system ; Wavelet Transform; MFCC ; Discrete to
Continuous algorithm .
I. INTRODUCTION
Identifying a speaker is one of the most complicated tasks
in speech recognition problems. It generally depends on the
productio n of speech of the speaker with physiological and
behavioral characteristics. These characteristics rely on the
speech generation (voice source) and the envelope behavior
(vocal and nasal tract) [1].
Speaker recognition can be arranged to speaker
identification and speaker verification [2]. Speaker
identification consists in comparing a vocal message with a set
of regi stered utterances corresponding to different speakers and
determines the one who spoke. It entails a multi -choice
classification problem. Speaker verification consists in
accepting or rejecting the speaker who claims to be. It entails a
yes-no hypothesis t esting problem.
All speaker recognition systems contain two main phases:
feature extraction and recognition. During the first one, a
training vector is generated from the speech signal of the word
(password) spoken by the user. These training vectors are
stored in a database for subsequent use in the recognition
phase. During the recognition phase, the system tries to
identify the unknown speaker by comparing extracted features
from password with the on es from a set of known speakers .
The reminder of this paper is organized as follows. Section
II presents conventional algorithms and the proposed system followed by section III that explains the processing applied on
spoken passwords. The procedure of feature extraction
especially MFCC is detailed in section IV. Section V explains
the matching algorithm. Section VI demonstrates the
experimental results followed by conclusion in section VII.
II. SPEAKER RECOGNITION SYSTEM
A. Conventional Algorithm
Speaker identification and speaker verification are two
aspects of spea ker recognition. Speaker identification
determines the person who spoke the given utterance am ongst
a given set of speakers. I n the other hand, speaker verification
process consists of accepting or rejecting the identity claimed
by a speaker. Speaker recog nition methods is divided into
text-dependent (fixed passwords) and text -independent (no
specified passwords) methods. Text -dependent systems
require less training than text – independent systems [3].
The identity of speaker can be represented by Linear
Predictive Cepstrum Coefficients (LPCC), delta LPCC,
MFCC, delta MFCC and pitch parameters [4]. Proso dy is also
a parameter that relies on the speaker characteristics [5]
Gaussian Mixture Model (GMM) is considered as robust
system for speaker recognition [6].
B. Developed System’s Architecture
The combination of algorithms and techniques improves
the accuracy or the recognition rate of many speech
applications.
In the proposed speaker recognition system, the signal is
preprocessed using endpoint detection algorithm and then we
used the discrete wave let transform to generate the
approximation coefficients of the speech signal, we notice that
we chose Daubechies (DAUB) family. In the features
extraction step, the MFCC method is applied to these
approximation coefficients to determine the speech feature s.
Finally, the input signal is compared with the stored models
using similarity measurements and then the most similar
model is chosen using the discrete to continuous algorithm.
Figure 1 shows the architecture of our proposed system based
on English pass word recognition .

Fig. 1. Speaker identification system architecture
III. SIGNAL PROCESSING
A. Signal Pre -processing
Speaker Pre-processing techniques of a signal are
extremely important for an optimal recognition. Speech
segments must be se parated from non -speech to achieve good
recognition accuracy in practical environment. The purpose of
the endpoint detection algorithm [7] is to determine the
beginning and the ending boundaries of each spoken word,
and also to delete the noise region and silence. The algorithm
relies on determining the energy level and zero -crossing rate.
Figure 2 shows the endpoint detection for spoken digit “1”
with its energy and zero -crossing rate.

Fig. 2. Result of Endpoint Detection fo r spoken digit “1” B. Wavelet Transform
Analyzing according to scale and translation is the prime
idea behind the wavelet transform [8]. Wavelet decomposition
satisfies certain mathematical r equirements which are used to
represent data or other functions. The Discrete Wavelet
Transform (DWT) is a case of the Wavelet Transform for
which the wavelet is sampled discretely. It provides a
compact representation of a signal in time and frequency . The
Discrete Wavelet Transform of a signal is computed by using
a number of filters (Low -pass and high -pass filters). The filter
outputs are then sub sampled by 2.
Different wavelet families have been presented such as
Haar wavlet, Morlet wavelet, Daubechie s (DAUB) wavelets,
etc. In our work, we employed the wavelet family proposed by
Daubechies [9].
IV. FEATURES EXTRACTION USING MFCC
Features extraction is the step of computing a sequence of
feature vectors that characterize the word. They provide a
comp act representation of the given speech signal. MFCC is
one of the most techniques used abundantly in automatic
speech recognition systems (ASR). It is widely used for both
speech and speaker recognition [10]. A Mel is a unit of
measure based on human ear’s perceived frequency. The Mel
scale is approximately linear spacing below 1000Hz and a
logarithmic spacing above 1000Hz. The Mel from frequency
can be ap proximately expressed as :

)700/ 1log( 2595)( f f mel  (1)
As illustrated in the block diagram below, a compact
representation would be provided by a set of MFCC, which are
the results of a cosine transform of the real logarithm of the
short -term energy spec trum expressed on a mel -frequency
scale.

Fig. 3. MFCC computation
The calculation of MFCC is obtained by following
contiguous steps:
-Pre-emphasis : The pre -emphasis filter is used to improve
the efficiency of the spectral analysis of the speech signal.
-Framing and windowing: The speech signal is usually
segmented to frames, and the spectral and cepstral analysis is
applied on each of these frames. Typically, the frame size is
25 milliseconds and the frames are overlapped by 10
milliseconds . After being separated into frames, each frame is Pre-
emphasis DFT
Output
energy
of filters
on mel-
scale Log Hamming
window Input
DCT MFCC database
Input
WaveletTransform
Matching
Decision Input MFCCs MFCCs model of speaker 1

MFCCs model of speaker N
MFCCs generation

multiplied by a window function before the spectral analysis
to reduce the discontinuity introduced by framing .
Windowing attenuates the values of the speech samples at the
beginning and the ending of each frame. Typically, Hamming
window is used.
-Spectral estimation: Fourier Transform (FFT) algorithm
is used to estimate the spectral coefficients of the speech
frames. Only the magnitude of the spectral coefficients is
retained and phase informati on is usually discarded .
-Mel filtering: To simul ate the characteristics of human’ s
ear, a set of triangular band pass filters are used to filter the
spectrum of speech signal. It is utilized to model the human
auditory system.
-Logarithmic c ompression : deals with the loudnes s non –
linearity. It approximately computes the relationship between
the human’ s perception of loudness and the sound intensity.
-Discrete Cosine Transform (DCT): The cepstrum is
defined as the inverse Fourier transform (IFFT) of the log
magnitude of Fourier transform of the signal. Since the log
Mel filter bank coefficients are real and symmetric, the IFFT
operation becomes DCT to determine the cepstral coeffic ients .
V. RECOGNITION USING SIMILARITY MEASUREMENTS
After a good feature extrac tion algorithm, the matching
stage is a primary task in almost speaker identification system
in term of reliability and also in term of time processing to
achieve the right decision. In matching phase, we chose to
implement the discrete to continuous algor ithm [11] as a
matcher algorithm. Where other algorithms tr y to find a good
matching of some pairs of features by performing a
comparison between the two features composition (point by
point), the discrete to continuous algorithm traits this problem
as point pattern matching problem. It brings the discrete
represe ntation of test signal onto th e continuous
representation (by interpolation) of the training dataset
considering the issue in its entity and not point by point as is
the case with the most algorithms dealing with this issue.
Let
nS and
mI represent the stored (training) template and
the input (testing) one respectively, where m and n denote the
number of features in S and I respectively. Both are extracted
using MFCC . In our case
mn
1) The problem is to decide whether or not there is an
allowed tran sformation
T , such that:
S IT)( . This problem
is knowing as finding the Largest Common
Substructures (LCS) .

The direct calculation of the transformatio n
Thas serious
difficulties if it is performed on a discrete representation
without prior knowledge of the correspondence. For this reason
the discrete to continuous algorithm use an intermediate step in
which the algorithm calculates a transformation
'T that brings
the discrete data
)(I on the continuous representation of the
model set
)(S .
Following , we describe the algorithm steps: 1. interpolation: Points (features) of
S interpolated using a
polynomial function P :
),(yx
i iy xP)(
(2)
Where
ix and
iy are the coordinates of
ith element of
S .
2. Search for
'T :
If the points of
I are included in
S , their transformation
after applying
'T must be included in the polynomial
representation of
S .
After transformation, we got
' )'(i i y xP where
'ix and
'iy
are the coordinates of
ith element of
I .
To calculate the
T' parameter the cost function
QT
is minimiz ed as follows:

 n
ij j y xP tyQT
12)" )"(( )(
(3)
Where
n is the number of the elements in
I .
We notice that the algorithm is implemented using JAVA
language programming.
VI. SIMULATION S AND RESULTS
In the proposed speech recogn ition system, for feature
extraction step, the MFCC are used to represent speech
features. The MFCC parameters are:
-window length: 0.025s
-step between successive windows: 0.016s
-number of cepstra to return MFCC , the first one
repre sents the en ergy: 13
-pre-emphasis: 0.97
45 speakers are tested. The password is spoken twice by
each speaker and total 90 utterances are collected in our
experiments. We have used 45 utterances for training and 45
utterances for testing. In all our experiments, the speech
signals are sampled at 8 kHz. We propose to identify speakers
based on their vocal password. The system differentiates
between speakers even if they use the same password “one”.
A. First Experiment
In thi s first experiment, we will measure the accuracy of
our proposed system applied only for 10 speakers. Each
speaker’s password is compared with the 10 speakers
(identification problem) .
After calculating the distances between the different
subjects , the minimal distance is used in decision step.

TABLE I. IDENTIFICATION O F SYSTEM ’S USERS
Speaker (test) Minimal
Distance Decision
Speaker 1 0.0010 0 Speaker 1
Speaker 2 0.000 49 Speaker 2
Speaker 3 0.000 43 Speaker 3
Speaker 4 0.000 37 Speaker 4
Speaker 5 0.000 89 Speaker 5
Speaker 6 0.000 37 Speaker 6
Speaker 7 0.000 22 Speaker 7
Speaker 8 0.000 43 Speaker 8
Speaker 9 0.000 33 Speaker 9
Speaker 10 0.000 82 Speaker 10
B. Second Experiment
In this scenario , we will test the thoroughness of the
developed algorithm for 45 speakers. Each speaker’s password
is compared with the 45 s peakers . The following table shows
the speakers that the system failed to identify.
TABLE II. CONFUSED USERS WITH T HEIR DISTANCES
Speaker Confused with Distance
Speaker 15 Speaker 26 0.0011
Speaker 19 Speaker 16 0.000975
Speaker 29 Speaker 5 0.00093
Total missed: 2 out of 45

The results illustrate that among 45 test samples; around
95% samples are correctly classified . To determine the
threshold so that we forbid a false acceptance, we must find
the minimum distance of false detection (it is equal to 0.0011
according to the table above). We can fix the threshold to
0.001 . In the case of the speaker 15 and regarding the used
thres hold (0.001 ), we can consider that the system didn’t fail
in the identification process becaus e the distance betwe en the
speakers 15 and 26 exceeds the th reshold. Thus, the speaker
will be claimed to spell the password again .
It is worth y to note that the recognition accuracy will be
improved when we use a remarkable number of trials in
training and testing phases.
VII. CONCLUSION
In this paper, we developed a method based on cepstral
analysis using MFCCs to extract speech features. We
proposed a word -dependent identification system where all
speakers use the same password. With the database used, the
recognition rate exceeded 95 %. MFCC method generates a
matrix of variable siz e depending on the number of frames.
Also the use of the wavelet transform has dramatically
enhanced the accuracy of the system. We believe that in
addition of the accuracy obtained by the system, the time
running is also optimized by using the mean value of MFCCs.
Also, more training data and good preprocessing methods can
increasingly enhance the accuracy of the system. REFERENCES
[1] D. Yu and L. Deng, Automatic speech recognition: a deep learning
appro ach. London: Springer, 2015.
[2] Y. K. Muthusamy, E. Barnard, and R. A. Cole, “Reviewing automatic
language identification,” IEEE Signal Processing Magazine , vol. 11, no.
4, pp. 33 –41, Oct. 1994.
[3] T. Kinnunen and H. Li, “An overview of text -independent speaker
recognition: From features to supervectors,” Speech Communication ,
vol. 52, no. 1, pp. 12 –40, Jan. 2010.
[4] Li Liu, Jialong He, and G. Palm, “Signal modeling for speaker
identification,” 1996, vol. 2, pp. 665 –668.
[5] M. J. Carey, E. S. Parris, H. Lloyd -Thomas, and S. Bennett, “Robust
prosodic features for speaker identification,” 1996, vol. 3, pp. 1800 –
1803.
[6] B. Xiang, U. V. Chaudhari, J. Navratil, G. N. Ramaswamy, and R. A.
Gopinath, “Short -time Gaussianization for robust speaker verification, ”
2002, p. I -681-I-684.
[7] L. R. Rabiner and M. R. Sambur, “An Algorithm for Determining the
Endpoints of Isolated Utterances,” Bell System Technical Journal , vol.
54, no. 2, pp. 297 –315, Feb. 1975.
[8] H.-G. Stark, Wavelets and Signal Processing: An Appl ication -Based
Introduction . Berlin, Heidelberg: Springer -Verlag Berlin Heidelberg,
2005.
[9] I. Daubechies, “Orthonormal bases of compactly supported wavelets,”
Communications on Pure and Applied Mathematics , vol. 41, no. 7, pp.
909–996, Oct. 1988.
[10] S. Engelberg, Digital signal processing: an experimental approach .
London: Springer, 2008.
[11] N. Aqili, M. Raji, A. Jilbab, S. Chaouki, and A. Hammouch, “PPM
Translation, Rotation and Scale in D -Dimensional Space by the Discrete
to Continuous Approach,” International Review on Computers and
Software (IRECOS) , vol. 11, no. 3, p. 270, Mar. 2016.

.