Performance Evaluation of Speaker Identification for Partial Coefficients of Transformed Full, Block and Row Mean of Speech Spectrogram using DCT, WALSH and HAAR

Report as inappropriate Your report has been received

Description

IJCSIS is an open access publishing venue for research in general computer science and information security. Target Audience: IT academics, university IT faculties; industry IT departments; government departments; the mobile industry and computing industry. Coverage includes: security infrastructures, network security: Internet security, content protection, cryptography, steganography and formal methods in information security; computer science, computer applications, multimedia systems, software, information systems, intelligent systems, web services, data mining, wireless communication, networking and technologies, innovation technology and management. The average paper acceptance rate for IJCSIS issues is kept at 25-30% with an aim to provide selective research work of quality in the areas of computer science and engineering. Thanks for your contributions in September 2010 issue and we are grateful to the experienced team of reviewers for providing valuable comments.

Shared by: ijcsis

Stats

views:: 121
posted:: 10/10/2010
language:: English
pages:: 13

Public Domain

Document Sample

Dr. H. B. Kekre et. al.(IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 6, 2010

Performance Evaluation of Speaker Identification for
Partial Coefficients of Transformed Full, Block and
Row Mean of Speech Spectrogram using DCT,
WALSH and HAAR
Dr. H. B. Kekre Dr. Tanuja K. Sarode Shachi J. Natu Prachi J. Natu
Senior Professor, Assistant Professor, Lecturer, Assistant Professor,
MPSTME, SVKM’s NMIMS Thadomal Shahani Engg. Thadomal Shahani Engg. GVAIET, Shelu
University College, Bandra (W), College, Bandra (W), Karjat 410201,
Mumbai, 400-056, India Mumbai, 400-050, India Mumbai, 400-050, India India
hbkekre@yahoo.com tanuja_0123@yahoo.com shachi_natu@yahoo.com prachi.natu@yahoo.com

Abstract- In this paper an attempt has been made to provide individual by these methods, he/she should be willing to
simple techniques for speaker identification using transforms undergo the tests and should not get upset by these procedures.
such as DCT, WALSH and HAAR alongwith the use of Speaker recognition allows non-intrusive monitoring and also
spectrograms instead of raw speech waves. Spectrograms form a achieves high accuracy rates which conform to most security
image database here. This image database is then subjected to requirements. Speaker recognition is the process of
different transformation techniques applied in different ways automatically recognizing who is speaking based on some
such as on full image, on image blocks and on Row Mean of an unique characteristics present in speaker’s voice [2]. There are
image and image blocks. In each method, results have been two major applications of speaker recognition technologies and
observed for partial feature vectors of image. From the results it methodologies: speaker identification and speaker verification.
has been observed that, transform on image block is better than
transform on full image in terms of identification rate and In the speaker identification task, a speech utterance from
computational complexity. Further, increase in identification rate an unknown speaker is analyzed and compared with speech
and decrease in computations has been observed when models of known speakers. The unknown speaker is identified
transforms are applied on Row Mean of an image and image as the speaker whose model best matches the input utterance.
blocks. Use of partial feature vector further reduces the number In speaker verification, an identity is claimed by an unknown
of comparisons needed for finding the most appropriate match. speaker, and an utterance of this unknown speaker is compared
with a model for the speaker whose identity is being claimed. If
Keywords- Speaker Identification, DCT, WALSH, HAAR, the match is good enough, that is, above a threshold, the
Image blocks, Row Mean, Partial feature vector. identity claim is accepted. The fundamental difference between
identification and verification is the number of decision
alternatives [3]. In identification, the number of decision
I. INTRODUCTION alternatives is equal to the size of the population, whereas in
To provide security in a multiuser environment, it has verification there are only two choices, acceptance or rejection,
become crucial to identify users and to grant access only to regardless of the population size. Therefore, speaker
those users who are authorized. Apart from the traditional login identification performance decreases as the size of the
and password method, use of biometric technology for the population increases, whereas speaker verification performance
authentication of users is becoming more and more popular approaches a constant, independent of the size of the
nowadays. Biometrics comprises methods for uniquely population, unless the distribution of physical characteristics of
recognizing humans based upon one or more intrinsic physical speakers is extremely biased.
or behavioral traits. Biometric characteristics can be divided in Speaker identification can be further categorized into text-
two main classes: Physiological which are related to the shape dependent and text independent speaker identification based on
of the body. Examples include fingerprint, face recognition, the relevance to speech contents [2, 4].
DNA, hand and palm geometry, iris recognition etc.
Behavioral, which are related to the behavior of a person. Text Dependent Speaker Identification requires the speaker
Examples include typing rhythm, gait and voice. Techniques saying exactly the enrolled or given password/speech. Text
like face recognition, fingerprint recognition and retinal blood Independent Speaker Identification is a process of verifying the
vessel patterns have their own drawbacks. To identify an identity without constraint on the speech content. It has no

186 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
Dr. H. B. Kekre et. al.(IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 6, 2010

advance knowledge of the speaker’s utterance and is more The MFCC parameter as proposed by Davis and
flexible in situation where the individuals submitting the Mermelstein [5] describes the energy distribution of speech
sample may be unaware of the collection or unwilling to signal in a frequency field. Wang Yutai et. al. [6] has proposed
cooperate, which presents more difficult challenge. a speaker recognition system based on dynamic MFCC
parameters. This technique combines the speaker information
Compared to Text Dependent Speaker Identification, Text obtained by MFCC with the pitch to dynamically construct a
Independent Speaker Identification is more convenient because set of the Mel-filters. These Mel-filters are further used to
the user can speak freely to the system. However, it requires extract the dynamic MFCC parameters which represent
longer training and testing utterances to achieve good
characteristics of speaker’s identity.
performance. Text Independent Speaker Identification is more
difficult problem as compared to Text Dependent Speaker Sleit, Serhan and Nemir [7] have proposed a histogram
Identification because the recognition system must be prepared based speaker identification technique which uses a reduced set
for an arbitrary input text. of features generated using MFCC method. For these features,
histograms are created using predefined interval length. These
Speaker Identification task can be further classified into histograms are generated first for all data in feature set for
closed set and open set identification. every speaker. In second approach, histograms are generated
In closed set problem, from N known speakers, the speaker for each feature column in feature set of each speaker.
whose reference template has the maximum degree of Another widely used method for feature extraction is use of
similarity with the template of input speech sample of unknown linear Prediction Coefficients (LPC). LPCs capture the
speaker is obtained. This unknown speaker is assumed to be information about short time spectral envelope of speech. LPCs
one of the given set of speakers. Thus in closed set problem, represent important speech characteristics such as formant
system makes a forced decision by choosing the best matching speech frequency and bandwidth [8].
speaker from the speaker database.
Vector Quantization (VQ) is yet another approach of
In the open set text dependent speaker identification, feature extraction [19-22]. In Vector Quantization based
matching reference template for an unknown speaker’s speech
speaker recognition systems; each speaker is characterized with
sample may not exist. So the system must have a predefined several prototypes known as code vectors [9]. Speaker
tolerance level such that the similarity degree between the recognition based on non-parametric vector quantization was
unknown speaker and the best matching speaker is within this proposed by Pati and Prasanna [10]. Speech is produced due to
tolerance. excitation of vocal tract. Therefore in this approach, excitation
In the proposed method, speaker identification is carried out information can be captured using LP analysis of speech signal
with spectrograms and transformation techniques such as DCT, and is called as LP residual. This LP residual is further
WALSH and HAAR [15-18]. Thus an attempt is made to subjected to non-parametric Vector Quantization to generate
formulate a digital signal processing problem into pattern codebooks of sufficiently large size. Combining nonparametric
recognition of images. Vector Quantization on excitation information with vocal tract
information obtained by MFCC was also introduced by them.
The rest of the paper is organized as follows: in section II
we present related work carried out in the field of speaker III. PROPOSED METHODS
identification. In section III our proposed approach is
In the proposed methods, first we converted the speech
presented. Section IV elaborates the experiment conducted and
samples collected from various speakers into spectrograms
results obtained. Analysis of computational complexity is
[11]. Spectrograms were created using Short Time Fourier
presented in section V. Conclusion has been outlined in section
Transfer method as discussed below:
VI.
In the approach using STFT, digitally sampled data are
II. RELATED WORK divided into chunks of specific size say 128, 256 etc. which
All speaker recognition systems at the highest level contain usually overlap. Fourier transform is then obtained to calculate
two modules, feature extraction and feature matching. the magnitude of the frequency spectrum for each chunk. Each
chunk then corresponds to a vertical line in the image, which is
Feature extraction is the process of extracting subset of a measurement of magnitude versus frequency for a specific
features from voice data that can later be used to identify the moment in time.
speaker. The basic idea behind the feature extraction is that the
entire feature set is not always necessary for the identification Thus we converted the speech database into image
process. Feature matching is the actual procedure of identifying database. Different transformation techniques such as Discrete
the speaker by comparing the extracted voice data with a Cosine Transform [12], WALSH transform and HAAR
database of known speakers and based on this suitable decision transform are then applied to these images in three different
is made. ways to obtain their feature vectors.
There are many techniques used to parametrically represent 1. Transform on full image
a voice signal for speaker recognition task. One of the most 2. Transform on image blocks obtained by dividing an
popular among them is Mel-Frequency Cepstrum Coefficients image into four equal and non-overlapping blocks
(MFCC) [1].

187 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
Dr. H. B. Kekre et. al.(IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 6, 2010

3. Transform on Row Mean of an image and on Row Step 5. Apply the transformation technique (DCT /
Mean of image blocks. WALSH / HAAR) on resized image to obtain its
feature vector.
From these feature vectors, again identification rate is
obtained for various portions selected from the feature vector Step 6. Save these feature vectors for further comparison.
i.e. for partial feature vector [15, 23, 24]. Two different sets of
database were generated. First set, containing 60% of the total Step 7. Calculate the Euclidean distance between feature
images as trainee images and 40% of the total images as test vectors of each test image with each trainee
images. Second set, containing 80% of the images as trainee image corresponding to the same sentence.
images and 20% of the total images as test images. Euclidean Step 8. Select the trainee image which has smallest
distance between test image and trainee image is used as a Euclidean distance with the test image and
measure of similarity. Euclidean distance between the points declare the speaker corresponding to this trainee
X(X1, X2, etc.) and point Y (Y1, Y2, etc.) is calculated using image as the identified speaker.
the formula shown in equation. (1).
Repeat Step 7 and Step 8 for partial feature vector
n
obtained from the full feature vector.
2
D= ∑ (X i − Yi ) (1) B. Transformation technique on image blocks[27, 29]:
i =1
In this second method, resized image of size 256*256 is
Smallest Euclidean distance between test image and trainee divided into four equal parts as shown in Fig. 2 and then 2-D
image means the most probable match of speaker. Algorithms DCT / WALSH / HAAR is applied to each part.
for transformation technique on full image and transformation
techniques on image blocks are given below. I II

A. Transformation techniques on full image[27, 28]: III IV

In the first method 2-D DCT / WALSH / HAAR is applied
on the full image resized to 256*256. Further, instead of full Fig. 2: Image divided into four equal non-overlapping blocks
feature vector of an image only some portion of feature vector
i.e. partial feature vector is selected for identification purpose. Thus when N*N image is divided into four equal and non-
This selection of feature vector is illustrated in Fig. 1 and it is overlapping blocks, blocks of size N/2*N/2 are obtained.
based on the number of rows and columns that have been Feature vector of each block when appended as columns forms
selected from the feature vector of an image. For example, a feature vector of an image. Thus size of feature vector of an
initially first full feature vector (i.e. 256*256) has been selected image in this case is of 128*512. Again Euclidean distance is
and then partial feature vectors of size 192*192, 128*128, used as a measure of similarity. Here also using partial feature
64*64, 32*32, 20*20 and 16*16 were selected from the feature vectors, identification rate has been obtained. Partial feature
vector. For these different sizes, identification rate was vectors of size 96*384, 64*256, 32*128, 16*64 and 8*32 have
obtained. been selected to find identification rate. Detailed steps are
explained in algorithm given below:
Step 1. For each trainee image in the database, resize an
image to size 256*256.
Step 2. Divide the image into four equal and non-
overlapping blocks as explained in Fig. 2.
Step 3. Apply transformation technique (DCT/ WALSH
/HAAR) on each block obtained in Step 2.
Fig. 1: Selection of partial feature vector Step 4. Append the feature vectors of each block one
after the other to get feature vector of an image.
Algorithm for this method is as follows:
Step 5. For each test image in the database, resize an
Step 1. For each trainee image in the database, resize an image to size 256*256.
image to size 256*256.
Step 6. Divide the image into four equal and non-
Step 2. Apply the transformation technique (DCT / overlapping blocks as shown in Fig. 2.
WALSH / HAAR) on resized image to obtain its
Step 7. Apply transformation technique (DCT /WALSH
feature vector.
/HAAR) on each block obtained in Step 6.
Step 3. Save these feature vectors for further comparison.
Step 8. Append the feature vectors of each block one
Step 4. For each test image in the database, resize an after the other to get feature vector of an image.
image to size 256*256.

188 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
Dr. H. B. Kekre et. al.(IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 6, 2010

Step 9. Calculate the Euclidean distance of each test IV. EXPERIMENTS AND RESULTS
image with each trainee image corresponding to Implementation for the proposed approach was done on
the same sentence. Intel Core 2 Duo Processor, 2.0 GHz, and 3 GB of RAM.
Step 10. Select the trainee image which has smallest Operating System used is Windows XP and softwares used are
Euclidean distance with the test image and MATLAB 7.0 and Sound forge 8.0. To study the proposed
declare the speaker corresponding to this trainee approach we recorded six distinct sentences from 30 speakers:
image as the identified speaker. 11 males and 19 females. These sentences are taken from
VidTIMIT database [13] and ELSDSR database [14]. For every
Repeat Step 9 and Step 10 for partial feature vectors speaker 10 occurrences of each sentence were recorded.
selected from feature vector obtained in Step 4 and Step 8. Recording was done at varying times. This forms the closed set
Selection of partial feature vector is similar to the one shown in for our experiment. From these speech samples spectrograms
Fig. 1. But in this method, size of feature vector is 128*512, were created with window size 256 and overlap of 128. Before
96*384, 64*256, 32*128, 16*64 and 8*32. creation of spectrograms, DC offset present in speech samples
C. Transformation techniques on Row Mean [16-18] of an was removed so that signals are vertically centered at 0. After
image and on Row Mean of image blocks [27, 29]: removal of DC offset, speech samples were normalized with
respect to amplitude to -3 dB and also with respect to time.
In this approach, Row Mean of an image is calculated. Row Spectrograms generated from these speech samples form the
mean is nothing but an average of pixel values of an image image database for our experiment. In all we had 1800
along each row. Fig. 3 shows how the Row Mean of an image spectrograms in our database.
is obtained.
From these spectrograms, two sets were created.
Row mean Set A: Contains six spectrograms as trainee images per
speaker and four spectrograms as test images per speaker. So in
all it contains 1080 trainee images and 720 test images.
Set B: Contains eight spectrograms as trainee images per
Fig. 3: Row Mean of an image
speaker and two spectrograms as test images per speaker. So in
all it contains 1440 trainee images and 360 test images.
1-D DCT / WALSH / HAAR is then applied on this Row
mean of an image to its feature vector and Euclidean distance is Since our work is restricted to text dependent approach,
used as measure of similarity to identify speaker. Detail Euclidean distance for a test image of speaker say ‘x’ for a
algorithm is given below: particular sentence say ‘s1’ is obtained by comparing the
feature vector of that test image with the feature vectors of all
Step 1: For each trainee image in the database, resize an the trainee images corresponding to sentence ‘s1’. Results are
image to size 256*256. calculated for set of test images corresponding to each
Step 2: Calculate Row Mean of an image as shown in sentence.
Fig. 3.
A. Results for DCT/WALSH/HAAR on Full image:
Step 3: Apply 1-D transformation technique (DCT /
WALSH / HAAR) on Row Mean obtained in 1) Results for DCT on full image
Step 2. This gives the feature vector of an image. Table I shows the identification rate for six sentences s1 to
Step 4: For each test image in the database, resize an s6 when DCT is applied on full image in set A and partial
image to size 256*256. feature vectors are selected to find the matching spectrogram.
Step 5: Apply 1-D transformation technique (DCT /
WALSH / HAAR) on Row Mean obtained in TABLE I. IDENTIFICATION RATE FOR SENTENCES S1 TO S6 FOR FULL AND
PARTIAL FEATURE VECTOR WHEN DCT IS APPLIED TO FULL IMAGE IN SET A
Step 4. This gives the feature vector of an image.
Step 6: Calculate the Euclidean distance of each test Portion of feature Sentence
image with each trainee image corresponding to vector selected S1 S2 S3 S4 S5 S6
the same sentence. 256*256 54.16 59.16 56.66 56.66 68.33 62.50
192*192 58.33 65 67.5 65 73.33 69.16
Step 7: Select the trainee image which has smallest 128*128 65.83 64.16 71.66 67.5 74.16 72.5
Euclidean distance with the test image and 64*64 70.83 70.83 71.66 72.50 77.50 75.83
declare the speaker corresponding to this trainee 32*32 75 73.33 74.16 75 80 77.5
image as the identified speaker. 20*20 78.33 75.33 78.33 71.66 81.66 80
For Row Mean of image blocks, first divide the image into 16*16 72.5 76.66 74.16 74.16 76.66 79.16
equal and non-overlapping blocks (of size 128*128, 64*64,
32*32, 16*16 and 8*8). Obtain Row Mean of each block as Table II shows the identification rate for six sentences s1 to
shown in Fig. 3. Transformation technique is then applied on s6 when DCT is applied on full image in set B and partial
Row Mean of each block and then combined into columns to feature vectors are selected to find the matching spectrogram.
get feature vector of an image.

189 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
Dr. H. B. Kekre et. al.(IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 6, 2010
TABLE II. IDENTIFICATION RATE FOR SENTENCES S1 TO S6 FOR VARYING TABLE V. IDENTIFICATION RATE FOR SENTENCES S1 TO S6 FOR FULL AND
PORTION OF FEATURE VECTOR WHEN DCT IS APPLIED TO FULL IMAGE IN SET B PARTIAL FEATURE VECTOR WHEN WALSH TRANSFORM IS APPLIED TO FULL
IMAGE FROM SET B
Portion of feature Sentence
vector selected S1 S2 S3 S4 S5 S6 Portion of feature Sentence
256*256 63.33 66.67 75 66.67 76.67 76.67 vector selected S1 S2 S3 S4 S5 S6
192*192 73.33 70 76.67 75 78.33 78.33 256*256 63.33 66.67 75 66.67 76.67 76.67
128*128 78.33 73.33 80 78.33 81.67 81.67 192*192 75 71.67 76.67 73.33 78.33 81.67
64*64 80 80 78.33 86.67 83.33 88.33 128*128 80 75 78.33 83.33 81.67 81.67
32*32 90 86.67 86.67 86.67 86.67 90 64*64 86.67 83.33 81.67 85 83.33 85
20*20 86.67 86.67 86.67 88.33 90 90 32*32 86.67 81.67 81.67 88.33 83.33 91.67
16*16 85 85 86.67 86.67 91.67 90 20*20 91.67 78.33 83.33 85 86.67 83.33
16*16 86.67 85 83.33 85 83.33 86.67

Table III shows the comparison of overall identification
Table VI shows the overall identification rate considering
rate considering all sentences, for partial feature vectors of
all sentences, for partial feature vectors. For set A, highest
different sizes when set A and set B is used. It also shows the
identification rate is obtained for partial feature vector of size
number of DCT coefficients used for identifying speaker for
64*64 i.e. 4096 WALSH coefficients. For set B, it requires
corresponding selected portion of feature vector.
32*32 partial feature vector i.e. 1024 WALSH coefficients.
TABLE III. COMPARISON OF OVERALL IDENTIFICATION RATE FOR
DIFFERENT NUMBER OF DCT COEFFICIENTS WHEN DCT IS APPLIED TO FULL TABLE VI. COMPARISON OF OVERALL IDENTIFICATION RATE FOR VARYING
IMAGE IN SET A AND SET B NUMBER OF COEFFICIENTS WHEN WALSH TRANSFORM IS APPLIED TO FULL
IMAGE FROM SET A AND SET B
% Identification rate
Portion of feature Number of DCT % Identification rate
Portion of feature Number of Walsh
vector selected coefficients Set A Set B vector selected coefficients Set A Set B
256*256 65536 60 70.83 256*256 65536 60 70.83
192*192 36864 66.38 75.27 192*192 36864 66.66 76.11
128*128 16384 69.30 78.88 128*128 16384 70.69 80
64*64 4096 73.19 82.77 64*64 4096 75 84.16
32*32 1024 75.83 87.77 32*32 1024 73.33 85.55
20*20 400 77.63 88.05 20*20 400 72.91 84.72
16*16 256 76.66 87.5 16*16 256 71.94 85

2) Results for Walsh on full image
3) Results for HAAR on full image
Results of Walsh transform on Spectrograms are tabulated
below. Table IV shows the identification rate for sentences s1 Table VII shows sentencewise identification rate when 2-D
to s6 for full and partial feature vectors when WALSH HAAR transform is applied to full image with size 256*256
transform is applied on full image and set A is used. and partial feature vectors are selected from these feature
vectors. These results are for set A.
TABLE IV IDENTIFICATION RATE FOR SENTENCES S1 TO S6 FOR VARYING
PORTION OF FEATURE VECTOR WHEN WALSH TRANSFORM IS APPLIED TO TABLE VII. IDENTIFICATION RATE FOR SENTENCES S1 TO S6 FOR VARYING
FULL IMAGE FROM SET A PORTION OF FEATURE VECTOR WHEN HAAR TRANSFORM IS APPLIED TO FULL
IMAGE FROM SET A
Portion of feature Sentence
vector selected S1 S2 S3 S4 S5 S6 Portion of feature Sentence
256*256 54.16 59.16 57.5 57.5 68.33 63.33 vector selected S1 S2 S3 S4 S5 S6
192*192 59.16 66.66 65.83 63.33 73.33 71.66 256*256 54.16 59.16 57.5 57.5 68.33 63.33
128*128 65.83 66.66 70.83 73.33 75 72.5 192*192 59.16 66.66 65.83 63.33 73.33 71.66
64*64 74.16 73.33 76.66 75 75.83 75 128*128 65.83 66.66 70.83 73.33 75 72.5
32*32 70.83 71.66 71.66 70.83 78.33 76.66 64*64 74.16 73.33 76.66 75 75.83 75
20*20 70.83 69.67 71.67 70.83 76.67 78.33 32*32 70.83 71.66 71.66 70.83 78.33 76.66
16*16 70 70.83 71.67 66.67 75 77.5 20*20 65.83 73.33 71.67 70 75 76.67
16*16 70 70.83 71.67 66.67 75 77.5

Table V shows sentencewise identification rate for
WALSH transform on full image from set B. It can be Table VIII shows identification rate for HAAR transform
observed from Table IV and Table V that, identification rate on full image when set B is used. From both the tables, it can
for each sentence is increased as more training is provided to be seen that as the number of coefficients selected from the
the system. From both the tables, it can be seen that as size of feature vector is decreased, the identification rate also
the partial feature vector is decreased, the identification rate decreases, achieves its peak value and then again decrease.
also decreases, achieves its peak value and then again decrease. From the Table VII and Table VIII, it can also be noted that
when more training is provided to the system, identification
rate per sentence is increased.

190 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
Dr. H. B. Kekre et. al.(IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 6, 2010
TABLE VIII. IDENTIFICATION RATE FOR SENTENCES S1 TO S6 FOR VARYING increases, reaches its maximum value and then again decreases
PORTION OF FEATURE VECTOR WHEN HAAR TRANSFORM IS APPLIED TO FULL
IMAGE WITH TRAINING SET OF EIGHT IMAGES FOR EACH SPEAKER
or remains constant.
Portion of feature Sentence TABLE XI. IDENTIFICATION RATE FOR SENTENCES S1 TO S6 FOR FULL AND
vector selected S1 S2 S3 S4 S5 S6 PARTIAL FEATURE VECTOR USING DCT ON IMAGE BLOCKS FOR IMAGES FROM
256*256 63.33 66.67 75 66.67 76.67 76.67 SET A
192*192 80 73.33 78.33 76.67 78.33 78.33
Portion of feature Sentence
128*128 80 75 78.33 83.33 81.67 81.67 vector selected S1 S2 S3 S4 S5 S6
64*64 86.67 83.33 81.67 85 83.33 85 128*512 54.16 59.16 57.5 57.5 68.33 63.33
32*32 86.67 81.67 81.67 88.33 83.33 91.67 96*384 60 63.33 65.33 65 73.33 68.33
20*20 86.67 88.33 86.67 85 85 86.67 64*256 65 65 70.83 66.66 74.16 71.16
16*16 86.67 85 83.33 85 83.33 86.67 32*128 70.83 70.83 70.83 71.66 76.66 75
16*64 75.83 74.16 75 75.83 81.66 77.5
Table IX shows identification rate obtained by considering 8*32 69.16 76.66 75 75.83 75 75.83
all six sentences, for set A and set B, with different sized partial
feature vectors. Maximum identification rate is observed for TABLE XII. IDENTIFICATION RATE FOR SENTENCES S1 TO S6 FOR FULL AND
PARTIAL FEATURE VECTOR USING DCT ON IMAGE BLOCKS FOR IMAGES FROM
4096 and 400 HAAR coefficients with set A and set B SET B
respectively.
Portion of feature Sentence
vector selected S1 S2 S3 S4 S5 S6
TABLE IX. COMPARISON OF OVERALL IDENTIFICATION RATE FOR VARYING
NUMBER OF COEFFICIENTS WHEN HAAR TRANSFORM IS APPLIED TO FULL 128*512 63.33 66.67 75 66.67 76.67 76.67
IMAGE FROM SET A AND SET B 96*384 71.67 70 76.67 75 78.33 78.33
64*256 78.33 73.33 80 76.67 81.67 81.67
Portion of feature Number of HAAR Identification rate (%) 32*128 78.33 80 78.33 86.67 83.33 86.67
vector selected coefficients Set A Set B 16*64 90 88.33 86.67 90 86.67 88.33
256*256 65536 60 70.83 8*32 88.33 88.33 85 86.67 90 86.67
192*192 36864 67.91 77.5
128*128 16384 70.69 80
64*64 4096 75 84.16 Table XIII shows the comparison of overall identification
32*32 1024 73.33 85.55 rate considering all sentences, for partial feature vectors using
20*20 400 72.08 86.39 DCT on image blocks. For both the training sets, maximum
16*16 256 71.94 85 identification rate is achieved for partial feature of size 16*64
i.e. for 1024 DCT coefficients.
Table X shows the comparison of identification rates for all
TABLE XIII. COMPARISON OF OVERALL IDENTIFICATION RATE FOR FULL AND
three transformation techniques on full image when set A and PARTIAL FEATURE VECTOR PORTION USING DCT ON IMAGE BLOCKS FOR
set B are used per speaker. IMAGES FROM SET A AND SET B

TABLE X. COMPARISON OF IDENTIFICATION RATES WHEN DCT, WALSH Portion of feature Number of DCT Identification rate (%)
AND HAAR ON FULL IMAGE FROM SET A AND SET B vector selected coefficients Set A Set B
128*512 65536 60 70.83
Portion of Identification rate (%) Identification rate (%)
96*384 36864 65.97 75
feature when set A is used when set B is used
vector 64*256 16384 68.88 78.61
DCT WALSH HAAR DCT WALSH HAAR 32*128 4096 72.63 82.22
selected
256*256 60 60 60 70.83 70.83 70.83 16*64 1024 76.66 88.33
8*32 256 74.58 86.67
192*192 66.38 66.66 67.91 75.27 76.11 77.5
128*128 69.30 70.69 70.69 78.88 80 80
64*64 73.19 75 75 82.77 84.16 84.16
32*32 75.83 73.33 73.33 87.77 85.55 85.55 2) Results for WALSH on image blocks:
20*20 77.63 72.91 72.08 88.05 84.72 86.39
16*16 76.66 71.94 71.94 87.5 85 85
Table IVX on the next page shows the sentencewise
identification rate when WALSH transform is applied to image
blocks using images in Set A.
B. Results for DCT/WALSH/HAAR on image block:
TABLE IVX. IDENTIFICATION RATE FOR SENTENCES S1 TO S6 FOR VARYING
1) Results for DCT on image blocks: PORTION OF FEATURE VECTOR USING WALSH ON IMAGE BLOCKS WITH
IMAGES FROM SET A
Table XI shows the identification rate for sentences s1 to s6
when full and partial feature vectors are selected to identify Portion of feature Sentence
speaker using DCT on image blocks using set A. Table XII vector selected S1 S2 S3 S4 S5 S6
shows the sentence wise identification rate for full and partial 128*512 54.16 59.17 57.5 57.5 68.33 63.33
96*384 59.17 66.67 65.83 63.33 73.33 71.67
feature vectors when DCT is applied on image blocks for set B. 64*256 65.83 66.67 70.83 73.33 75 72.5
It can be seen from the table that identification rate is improved 32*128 74.17 73.33 76.67 75 75.83 75
when more training is provided to the system. From both the 16*64 70.83 71.67 71.67 70.83 78.33 76.67
tables, it can be seen that, as the number of coefficients used 8*32 70 70.83 71.67 66.67 75 77.5
for identification purpose decreases, the identification rate

191 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
Dr. H. B. Kekre et. al.(IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 6, 2010

Table XV show the sentencewise identification rate when used for identification purpose decreases, the identification rate
WALSH transform is applied to image blocks using Set B. increases, reaches some peak value and then again decreases or
From Table IVX and Table XV, it can be seen that, as the remains constant.
number of coefficients used for identification purpose
decreases, the identification rate increases, reaches its TABLE XVIII. IDENTIFICATION RATE FOR SENTENCES S1 TO S6 FOR FULL AND
PARTIAL FEATURE VECTOR USING HAAR ON IMAGE BLOCKS USING SET B
maximum value and then again decreases or remains constant.
Table XVI summarizes overall identification rate for both Portion of Sentence
training sets for various partial feature vectors. feature vector
S1 S2 S3 S4 S5 S6
selected
TABLE XV. IDENTIFICATION RATE FOR SENTENCES S1 TO S6 FOR VARYING 128*512 63.33 66.67 75 66.67 76.67 76.67
PORTION OF FEATURE VECTOR USING WALSH ON IMAGE BLOCKS WITH 96*384 73.33 70 78.33 76.67 78.33 81.67
IMAGES FROM SET B 64*256 80 75 78.33 83.33 81.67 81.67
32*128 86.67 83.33 81.67 85 83.33 85
Portion of feature Sentence 16*64 86.67 81.67 81.67 88.33 83.33 91.67
vector selected S1 S2 S3 S4 S5 S6 8*32 86.67 85 83.33 85 83.33 86.67
128*512 63.33 66.67 75 66.67 76.67 76.67
96*384 75 71.67 76.67 73.33 78.33 81.67
64*256 80 75 78.33 83.33 81.67 81.67 Table XIX shows overall identification rate for both
32*128 86.67 83.33 81.67 85 83.33 85 training sets obtained by considering the identification rate for
16*64 86.67 81.67 81.67 88.33 83.33 91.67 each sentence for various partial feature vectors. For Set A, the
8*32 86.67 85 83.33 85 83.33 86.67 maximum identification rate of 75.27% is obtained for 32*128
feature vector. Whereas, for Set B, the maximum identification
TABLE XVI. COMPARISON OF OVERALL IDENTIFICATION RATE FOR VARYING
SIZE OF FEATURE VECTOR PORTION USING WALSH ON IMAGE BLOCKS FOR
rate of 85.55% is obtained for 16*64 feature vector. Table XX
IMAGES IN SET A AND SET B shows comparison of overall identification rates for all three
transformation techniques when applied on image blocks for
Portion of Number of Identification rate (%) Set A and Set B.
feature vector WALSH
Set A Set B
selected coefficients
128*512 65536 60 70.83 TABLE XIX. COMPARISON OF OVERALL IDENTIFICATION RATE FOR VARYING
96*384 36864 66.67 76.11 SIZE OF FEATURE VECTOR PORTION USING HAAR ON IMAGE BLOCKS USING
SET A AND SET B
64*256 16384 70.69 80
32*128 4096 75 84.16 Portion of feature Number of HAAR Identification rate (%)
16*64 1024 73.33 85.55 vector selected coefficients Set A Set B
8*32 256 71.94 85 128*512 65536 59.86 70.83
96*384 36864 65.97 76.39
It can be observed from Table XVI that the maximum 64*256 16384 70.69 80
identification rate in case of Set A is obtained for 4096 32*128 4096 75.27 84.44
16*64 1024 73.33 85.55
WALSH coefficients i.e. for partial feature vector of size 8*32 256 71.94 85.27
32*128. The maximum identification rate in case of Set B is
obtained for 1024 WALSH coefficients i.e. for partial feature TABLE XX. COMPARISON OF IDENTIFICATION RATES W HEN DCT, WALSH
vector of size 16*64. AND HAAR ARE APPLIED ON IMAGE BLOCKS FOR IMAGES IN SET A AND SET B

3) Results for HAAR on image blocks: Portion of Identification rate (%) Identification rate (%)
feature When Set A is used When Set B is used
Table XVII shows identification rate for each sentence vector
when 2-D HAAR transform is applied on image blocks selected DCT WALSH HAAR DCT WALSH HAAR
obtained by dividing an image into four equal and non- 128*512 60 60 59.86 70.83 70.83 70.83
overlapping blocks as shown in Fig. 2. These results are for 96*384 65.97 66.67 65.97 75 76.11 76.39
training Set A. 64*256 68.88 70.69 70.69 78.61 80 80
32*128 72.63 75 75.27 82.22 84.16 84.44
TABLE XVII. IDENTIFICATION RATE FOR SENTENCES S1 TO S6 FOR VARYING 16*64 76.66 73.33 73.33 88.33 85.55 85.55
PORTION OF FEATURE VECTOR USING HAAR ON IMAGE BLOCKS USING SET A 8*32 74.58 71.94 71.94 86.67 85 85.27

Portion of feature Sentence
vector selected S1 S2 S3 S4 S5 S6 C. Results for DCT/ WALSH/ HAAR on Row Mean of an
128*512 53.33 59.17 57.5 57.5 68.33 63.33 image and Row Mean of image blocks :
96*384 58.33 61.67 65.83 68.33 72.5 69.17
64*256 65.83 66.67 70.83 73.33 75 72.5 1) Results for DCT on Row Mean of an image :
32*128 74.17 73.33 76.67 76.67 75 75.83
16*64 70.83 71.67 71.67 70.83 78.33 76.67 Table XXI shows sentence wise results obtained for Set A
8*32 70 70.83 71.67 66.67 75 77.5 when DCT of Row Mean is taken by dividing an image into
different number of non-overlapping blocks.
Table XVIII shows results when Set B is used and 2-D
HAAR transform is applied on image blocks. For both the
training sets, it is observed that, as the number of coefficients

192 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
Dr. H. B. Kekre et. al.(IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 6, 2010
TABLE XXI. IDENTIFICATION RATE FOR SENTENCES S1 TO S6 FOR DCT ON when it is divided into different number of non-overlapping
ROW MEAN OF AN IMAGE WHEN IMAGE IS DIVIDED INTO DIFFERENT NUMBER
OF NON-OVERLAPPING BLOCKS USING SET A
and Set A is used. Table XXV shows the sentence wise
identification rate when Walsh transform is applied to Row
No. of blocks for Sentence Mean of an image when it is divided into different number of
image split S1 S2 S3 S4 S5 S6 non-overlapping and Set B.
Full image
57.5 66.66 64.16 60.83 60.83 62.5
(256*256) TABLE IVXX. IDENTIFICATION RATE FOR SENTENCES S1 TO S6 FOR WALSH
4 Blocks TRANSFORM ON ROW MEAN OF AN IMAGE WHEN IMAGE IS DIVIDED INTO
60.83 70.83 63.33 65.83 70 65.83 DIFFERENT NUMBER OF NON-OVERLAPPING BLOCKS WITH SET A
(128*128)
16 Blocks
69.16 75.83 70.83 65.83 73.33 71.66
(64*64) No. of blocks for image Sentence
64 Blocks split S1 S2 S3 S4 S5 S6
75 76.66 75.83 70 78.83 75.83
(32*32) Full image (256*256) 57.5 66.66 64.16 60.83 60.83 62.5
256 Blocks 4 Blocks (128*128) 60.83 70.83 63.33 65.83 70 65.83
76.66 75 75.83 72.5 80 82.5
(16*16) 16 Blocks (64*64) 69.16 75.83 70.83 65.83 73.33 71.66
1024 Blocks 64 Blocks (32*32) 75 76.66 75.83 70 78.83 75.83
74.16 72.5 75 72.5 80.83 78.33
(8*8) 256 Blocks (16*16) 76.66 75 75.83 72.5 80 82.5
1024 Blocks (8*8) 74.16 72.5 75 72.5 80.83 78.33

It can be seen from the Table XXI that, as the block size
chosen for calculating Row Mean reduces, better identification TABLE XXV. IDENTIFICATION RATE FOR SENTENCES S1 TO S6 FOR WALSH
rate is achieved. For block size 16*16, maximum identification TRANSFORM ON ROW MEAN OF AN IMAGE WHEN IMAGE IS DIVIDED INTO
rate is obtained and then it decreases again. DIFFERENT NUMBER OF NON-OVERLAPPING BLOCKS WITH SET B

Table XXII shows the identification rate for sentence s1 to
No. of blocks for image Sentence
s6 and Set B of images. It can be seen from the Table XXII split S1 S2 S3 S4 S5 S6
that, as the block size chosen for calculating Row Mean Full image (256*256) 73.33 76.66 78.33 76.66 75 80
reduces, better identification rate is achieved. For block size 4 Blocks (128*128) 80 80 78.33 81.67 81.67 80
16*16, maximum identification rate is obtained and then it 16 Blocks (64*64) 91.67 81.67 83.33 83.33 81.67 83.33
decreases again. 64 Blocks (32*32) 91.67 85 86.66 86.66 86.66 88.33
256 Blocks (16*16) 91.67 88.33 88.33 85 91.67 90
TABLE XXII. IDENTIFICATION RATE FOR SENTENCES S1 TO S6 FOR DCT ON 1024 Blocks (8*8) 88.33 83.33 85 86.66 85 88.33
ROW MEAN OF AN IMAGE WHEN IMAGE IS DIVIDED INTO DIFFERENT NUMBER
OF NON-OVERLAPPING BLOCKS USING SET B
It can be seen from the Table IVXX and Table XXV that,
No. of blocks for image Sentence as the block size chosen for calculating Row Mean reduces,
split S1 S2 S3 S4 S5 S6 better identification rate is achieved.
Full image (256*256) 73.33 76.67 78.33 76.67 75 80
4 Blocks (128*128) 80 80 78.33 81.67 81.67 80 Table XXVI summarizes overall identification rate for both
16 Blocks (64*64) 91.67 81.67 83.33 83.33 81.67 83.33 training sets by considering all six sentences. For block size
64 Blocks (32*32) 91.67 85 86.67 86.67 86.67 88.33 16*16, maximum identification rate is obtained and then it
256 Blocks (16*16) 91.67 88.33 88.33 85 91.67 90 decreases again.
1024 Blocks (8*8) 88.33 83.33 85 86.67 85 88.33
TABLE XXVI. COMPARISON OF OVERALL IDENTIFICATION RATE FOR WALSH
The overall identification rates for both the sets, when DCT TRANSFORM ON ROW MEAN OF AN IMAGE WHEN IMAGE IS DIVIDED INTO
DIFFERENT NUMBER OF NON-OVERLAPPING BLOCKS FOR SET A AND SET B
of Row Mean is taken by dividing an image into different
number of non-overlapping blocks are tabulated in Table Identification rate
No. of blocks for Number of Walsh (%)
XXIII. image split coefficients
Set A Set B
TABLE XXIII. COMPARISON OF OVERALL IDENTIFICATION RATE FOR DCT ON Full image (256*256) 256 62.08 76.67
ROW MEAN OF AN IMAGE WHEN IMAGE IS DIVIDED INTO DIFFERENT NUMBER 4 Blocks (128*128) 512 66.11 80.27
OF NON-OVERLAPPING BLOCKS WITH SET A AND SET B 16 Blocks (64*64) 1024 71.11 84.17
Number of Identification rate 64 Blocks (32*32) 2048 75.27 87.5
No. of blocks for image 256 Blocks (16*16) 4096 77.08 89.17
DCT (%)
split 1024 Blocks (8*8) 8192 75.55 86.11
coefficients For Set A For Set B
Full image (256*256) 256 62.08 76.67
4 Blocks 128*128) 512 66.11 80.27
16 Blocks (64*64) 1024 71.11 84.17 3) Results for HAAR on Row Mean of an image :
64 Blocks (32*32) 2048 75.27 87.5 Table XXVII shows identification rate for each sentence
256 Blocks 16*16) 4096 77.08 89.17 when 1-D HAAR transform is applied to Row Mean of an
1024 Blocks (8*8) 8192 75.55 86.11 256*256 image and when image is divided into different
number of non-overlapping blocks for Set A. Similarly, Table
2) Results forWALSH on Row Mean of an image : XXVIII shows identification rate for each sentence when 1-D
HAAR transform is applied to Row Mean of an 256*256 image
Table IVXX shows the sentence wise identification rate
when Walsh transform is applied to Row Mean of an image

193 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
Dr. H. B. Kekre et. al.(IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 6, 2010

and when image is divided into different number of non- TABLE XXX. COMPARISON OF DCT, WALSH AND HAAR ON ROW
MEAN OF IMAGE AND IMAGE BLOCKS
overlapping blocks for Set B.
TABLE XXVII. IDENTIFICATION RATE FOR SENTENCES S1 TO S6 FOR No. of Identification rate (%) Identification rate (%)
HAAR TRANSFORM ON ROW MEAN OF AN IMAGE WHEN IMAGE IS DIVIDED blocks for when Set A is used when Set B is used
INTO DIFFERENT NUMBER OF NON-OVERLAPPING BLOCKS USING SET A image
split DCT WALSH HAAR DCT WALSH HAAR
Full image
No. of blocks for image Sentence 62.08 62.08 62.08 76.67 76.67 76.67
(256*256)
split S1 S2 S3 S4 S5 S6
4 Blocks
Full image (256*256) 57.5 66.66 64.16 60.83 60.83 62.5 66.11 66.11 66.11 80.27 80.27 80.27
(128*128)
4 Blocks (128*128) 60.83 70.83 63.33 65.83 70 65.83
16 Blocks
16 Blocks (64*64) 69.16 75.83 70.83 65.83 73.33 71.66 71.11 71.11 71.11 84.17 84.17 84.17
(64*64)
64 Blocks (32*32) 75 76.66 75.83 70 78.83 75.83
64 Blocks
256 Blocks (16*16) 76.66 75 75.83 72.5 80 82.5 75.27 75.27 75.27 87.5 87.5 87.5
(32*32)
1024 Blocks (8*8) 74.16 72.5 75 72.5 80.83 78.33 256
Blocks 77.08 77.08 77.08 89.17 89.17 89.17
TABLE XXVIII. IDENTIFICATION RATE FOR SENTENCES S1 TO S6 FOR (16*16)
HAAR TRANSFORM ON ROW MEAN OF AN IMAGE WHEN IMAGE IS DIVIDED 1024
INTO DIFFERENT NUMBER OF NON-OVERLAPPING BLOCKS USING SET B Blocks 75.55 75.55 75.55 86.11 86.11 86.11
(8*8)

No. of blocks for Sentence
image split S1 S2 S3 S4 S5 S6 V. COMPLEXITY ANALYSIS
Full image
73.33 76.67 78.33 76.67 75 80 A. Complexity analysis of DCT, WALSH and HAAR on full
(256*256)
4 Blocks image:
80 80 78.33 81.67 81.67 80
(128*128) For 2-D DCT on N*N image, 2N3 multiplications are
16 Blocks
91.67 81.67 83.33 83.33 81.67 83.33 required and 2N2(N-1) additions are required. For 2-D WALSH
(64*64)
on N*N image, 2N2(N-1) additions are required. For 2-D
64 Blocks
(32*32)
91.67 85 86.67 86.67 86.67 88.33 HAAR transform on N*N image where N=2m, number of
256 Blocks multiplications required are 2(m+1)N2 and number of additions
91.67 88.33 88.33 85 91.67 90
(16*16) required are 2mN2. Table XXXI summarizes these details
1024 Blocks
88.33 83.33 85 86.67 85 88.33 along with actual values of mathematical computations needed
(8*8) for processing of 256*256 images.
TABLE XXXI. COMPARISON BETWEEN DCT, WALSH AND HAAR
Table XXIX shows overall identification rate for the two WITH RESPECT TO MATHEMATICAL COMPUTATIONS AND IDENTIFICATION RATE
WHEN APPLIED ON FULL IMAGE
training sets when 1-D HAAR transform is applied to an image
divided into different number of equal and non-overlapping Algorithm
blocks. HAAR on
Parameter DCT on full WALSH on full
full image
TABLE XXIX. COMPARISON OF OVERALL IDENTIFICATION RATE FOR image(N*N) image(N*N)
(N*N)
HAAR TRANSFORM ON ROW MEAN OF AN IMAGE WHEN IMAGE IS DIVIDED Number of
INTO DIFFERENT NUMBER OF NON-OVERLAPPING BLOCKS FOR SET A AND B. 2N3 0 2(m+1)N2
Multiplications
Number of Identification rate (%) N=256 33554432 0 1179648
No. of blocks for Number of
HAAR 2N2(N-1) 2N2(N-1) 2mN2
image split For Set A For Set B Additions
coefficients
Full image N=256 33423360 33423360 1048576
256 62.08 76.67 Identification rate
(256*256) 77.63 75 75
4 Blocks (%) for Set A
512 66.11 80.27 Identification rate
(128*128) 88.05 85.55 86.39
16 Blocks (%) for Set B
1024 71.11 84.17
(64*64)
64 Blocks
2048 75.27 87.5 From the above table, it can be seen that DCT on full image
(32*32)
256 Blocks gives the highest identification rate for both the training sets as
4096 77.08 89.17
(16*16) compared to WALSH and HAAR on full image. However this
1024 Blocks outstanding performance is achieved at the expense of higher
8192 75.55 86.11
(8*8)
computations.

Overall identification rate for DCT, WALSH and HAAR Number of multiplications required by DCT on full image
on Row Mean of an image and image blocks are summarized in is approximately 28 times more than the number of
the Table XXX. multiplications required by HAAR on full image. Whereas
number of additions required by DCT on full image is
approximately 31 times more than the number of additions
required by HAAR on full image.

194 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
Dr. H. B. Kekre et. al.(IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 6, 2010

Though WALSH on full image does not require any additions are required. One dimensional Walsh on Row Mean
multiplications, overall CPU time taken by it is more than that of an image takes N(N-1) additions and no multiplications.
of HAAR on full image. This is because the number of Whereas, 1-D HAAR on Row Mean of an image of size N*N
additions taken by WALSH on full image is approximately 31 requires (m+1)N multiplications and mN additions where
times more than the number of additions required by HAAR on N=2m. Following Table XXXIII summarizes this statistics in
full image. case of each transformation technique applied for the Row
Mean of block size 16*16 which gives highest identification
B. Complexity analysis of DCT, WALSH and HAAR on rate.
image blocks:
TABLE XXXIII. COMPARISON BETWEEN DCT, WALSH AND HAAR
The number of multiplications required in case of 2-D DCT WITH RESPECT TO MATHEMATICAL COMPUTATIONS AND IDENTIFICATION RATE
on image blocks is N3 and the number of additions required are WHEN APPLIED ON ROW MEAN OF AN IMAGE
N2(N-2).
Algorithm
For 2-D WALSH on four image blocks of size N/2*N/2, Walsh on
DCT on HAAR on
number of additions required are N2(N-2). Row
Parameter Row Mean Row Mean
Mean of
of image of image
The number of multiplications required for 2-D HAAR on image
(N*1) (N*1)
image blocks is 2mN2. Similarly number of additions required (N*1)
for 2-D HAAR on image blocks is 2(m-1)N2. Table XXXII Number of
N2 0 (m+1)N
summarizes these details along with actual values of Multiplications
mathematical computations needed for processing of 256*256
N=16, 256 blocks 65536 0 20480
images.
TABLE XXXII COMPARISON BETWEEN DCT, WALSH AND HAAR WITH Number of Additions N(N-1) N(N-1) mN
RESPECT TO MATHEMATICAL COMPUTATIONS AND IDENTIFICATION RATE
WHEN APPLIED ON IMAGE BLOCKS N=16, 256 blocks 61440 61440 16384

Algorithm Identification rate (%)
77.08 77.08 77.08
Walsh on HAAR on for Set A
DCT on
Parameter image image
image blocks Identification rate (%)
blocks blocks 89.17 89.17 89.17
(N/2*N/2) for Set B
(N/2*N/2) (N/2*N/2)
Number of
N3 0 2(m+1)N2
Multiplications
From the Table XXXIII, we can see that all three
N=256, four blocks 16777216 0 1048576 transformation techniques result in same identification rate
Number of Additions N2(N-2) N2(N-2) 2(m-1)N2
when applied on the Row Mean of an image and on Row Mean
of an image blocks. For both the training sets, highest
N=256, four blocks 16646144 16646144 917504 identification rate is obtained when image is divided into 16*16
Identification rate (%) size blocks. However, in terms of computations, HAAR
76.66 75 75.27 transform is proved to be better one. Number of multiplications
for Set A
Identification rate (%) required by HAAR is approximately three times less than the
88.33 85.55 85.55
for Set B number of multiplications required in case of DCT on Row
Mean. Also the number of additions required by HAAR is 3.5
From the Table XXXII, it can be seen that, in all the three times less than the number of additions required by DCT and
transformation techniques on image blocks, DCT on image WALSH on Row Mean of an image.
blocks gives best identification rate for both the training sets. Along with the different approaches of applying
But this performance is achieved at the expense of higher transformation techniques on spectrograms, comparative study
number of computations. DCT on image blocks takes 16 times of computational complexity of three transformation
more multiplications and 18 times more additions than HAAR techniques for each approach has been done and is presented
on image blocks. Though WALSH transform does not need below.
any multiplications, still it takes more number of computations
than HAAR. This is because WALSH on image blocks requires D. Complexity Analysis of DCT transform on Full, Block and
approximately 18 times more additions than HAAR on image Row Mean of Spectrograms:
blocks. For 2-D DCT on N*N image, 2N3 multiplications are
required and 2N2(N-1) additions are required. For 2-D DCT on
C. Complexity Analysis of DCT, Walsh and HAAR on Row four blocks of size N/2*N/2, N3 multiplications are required
Mean of an image and on Row Mean of image blocks: and N2(N-2) additions are required. For 1-D DCT on N*1
Since Row Mean of an image is a one dimensional vector, image, N2 multiplications are needed and N(N-1) additions are
only 1-D DCT, WALSH and HAAR need to be applied on needed. These computational details are summarized in Table
Row Mean. This itself reduces the number of multiplications XXXIV along with the actual number of computations for
and additions required for feature vector calculation. Row 256*256 image using three methods of applying DCT.
Mean of an image of size N*N is a vector of size N*1. For 1-D
DCT on this N*1 vector, N2 multiplications and N(N-1)

195 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
Dr. H. B. Kekre et. al.(IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 6, 2010
TABLE XXXIV. COMPUTATIONAL DETAILS FOR 2-D DCT ON N*N WALSH transform is applied to Row Mean of an image. Also
IMAGE, 2-D DCT ON N/2*N/2 IMAGE AND 1-DCT ON N*1 IMAGE
RESPECTIVELY
the number of additions required when WALSH transform is
applied on image blocks is 255 times more than the number of
Parameter → No. of additions required when WALSH transform is applied to Row
No. of Additions
Algorithm ↓ Multiplications Mean of an image. Thus number of additions is drastically
2-D DCT on N*N image 2N3 2N2(N-1) reduced for Walsh transform on Row Mean of an image.
2-D DCT on 256*256
33554832 33424159
image F. Complexity Analysis of HAAR transform on Full, Block
2-D DCT on four blocks of N 3
N2(N-2) and Row Mean of Spectrograms:
size N/2*N/2 each
2-D DCT on four blocks of For 2-D HAAR transform on N*N image where N=2m,
size 256/2*256/2 each
16778240 16648191 number of multiplications required are 2(m+1)N2 and number
1-D DCT on N*1 Row
N2 N(N-1)
of additions required are 2mN2. For 2-D HAAR transform on
Mean of N*N image four blocks of size N/2*N/2 each, 2mN2 multiplications and
1-D DCT on N*1 Row 2(m-1)N2 additions are needed. Whereas for 1-D HAAR
69632 69631
Mean of 256*1 image
transform on N*1 image, number of multiplications required
are (m+1)N and number of additions are mN as shown in table
When all three methods of applying DCT are compared, it XXXVI.
has been observed that though number of coefficients used in
Row Mean method is higher, number of multiplications and TABLE XXXVI. COMPUTATIONAL DETAILS FOR 2-D HAAR ON N*N
IMAGE, 2-D HAAR ON N/2*N/2 IMAGE AND 1-D HAAR ON N*1 IMAGE
additions reduce drastically as compared to other two methods. RESPECTIVELY
Number of multiplications in DCT on full image method is 480
times more than the number of multiplications in Row Mean Parameter → No. of
No. of Additions
method whereas for DCT on image blocks it is 241 times more. Algorithm ↓ Multiplications
2-D HAAR on N*N image 2(m+1)N2 2mN2
Number of additions needed in DCT on full image and DCT on
2-D HAAR on 256*256
image blocks are also 480 times and 239 times more than the image
1179648 1048576
additions required in Row mean method respectively. For the 2-D HAAR on four blocks
Set A, the identification rate for DCT on Row Mean is almost 2mN2 2(m-1)N2
of size N/2*N/2 each
same as identification rate for DCT on full image. In case of 2-D HAAR on four blocks
1048576 917504
Set B, DCT on Row Mean gives better identification rate as of size 256/2*256/2 each
compared to DCT on full image and DCT on image blocks and 1-D HAAR on N*1 image (m+1)N mN
1-D HAAR on 256*1 size
that too with reduced number of mathematical computations. Row Mean vector of 2304 2048
E. Complexity Analysis of WALSH transform on Full, Block image 256*1
and Row Mean of Spectrograms:
For 2-D WALSH on N*N image, 2N2(N-1) additions are From Table XXXVI, it can be seen that number of
required. For 2-D WALSH on four blocks of size N/2*N/2, multiplications required when HAAR transform is applied on
N2(N-1) additions are required. Whereas for 1-D WALSH on full image is 512 times more than the number of
N*1 image, N(N-1) additions are needed as shown in table multiplications required when HAAR transform is applied to
XXXV. In all three cases number of multiplications required is Row Mean of an image. Also the number of multiplications
zero. required when HAAR transform is applied on image blocks is
455 times more than the number of multiplications required
TABLE XXXV. COMPUTATIONAL DETAILS FOR 2-D WALSH ON N*N
IMAGE, 2-D WALSH ON N/2*N/2 IMAGE AND 1-D WALSH ON N*1 IMAGE when HAAR transform is applied to Row Mean of an image.
RESPECTIVELY Thus number of multiplications is drastically reduced for
HAAR transform on Row Mean of an image. Number of
Parameter → No. of No. of
Multiplications Additions
additions required is also reduced to a greater extent when
Algorithm ↓
transformation technique is applied on Row Mean of an image.
2-D WALSH on N*N image 0 2N2(N-1)
Number of additions required when HAAR transform is
2-D WALSH on 256*256 applied on full image is 512 times more than the number of
0 33423360
image
2-D WALSH on four blocks of
additions required when HAAR transform is applied to Row
0 N2(N-2) Mean of an image. Also the number of additions required when
size N/2*N/2 each
2-D WALSH on four blocks of HAAR transform is applied on image blocks is 448 times more
0 16646144
size 256/2*256/2 each than the number of additions required when HAAR transform
1-D WALSH on N*1 size Row is applied to Row Mean of an image.
0 N(N-1)
Mean vector of image N*N
1-D WALSH on 256*1 size VI. CONCLUSION
Row Mean vector of image 0 65280
256*1 In this paper, closed set text dependent speaker
identification has been considered using three different
transformation techniques namely DCT, WALSH and HAAR.
From table 7.5 it can be seen that number of additions Each transformation technique is applied in three ways:
required when WALSH transform is applied on full image is
512 times more than the number of additions required when a) On full image

196 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
Dr. H. B. Kekre et. al.(IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 6, 2010

b) On image blocks and compared to transformation techniques on full image and on
image blocks. HAAR transform on Row Mean of an image
c) On Row Mean of an image.
gives the best result with respect to identification rate as well as
For each method, two training sets were used as mentioned number of computations required.
earlier.
REFERENCES
It can be clearly concluded from the results that as more [1] Evgeniy Gabrilovich, Alberto D. Berstin: “Speaker recognition: using a
training is provided to the system, more accuracy is obtained in vector quantization approach for robust text-independent speaker
the results in terms of identification rate. identification”, Technical report DSPG-95-9-001’, September 1995.
[2] Tridibesh Dutta, “Text dependent speaker identification based on
Further for each method, Identification rates are obtained spectrograms”, Proceedings of Image and vision computing, pp. 238-
for various numbers of coefficients from feature vectors of 243, New Zealand 2007,.
images. It has been observed that as the number of coefficients [3] J.P.Campbell, “Speaker recognition: a tutorial”, Proc. IEEE, vol. 85, no.
chosen is smaller up to a certain limit; better identification rate 9, pp. 1437-1462, 1997.
is achieved in all three methods. [4] D. O’Shaughnessy, “Speech communications- Man and Machine”, New
York, IEEE Press, 2nd Ed., pp. 199, pp. 437-458, 2000.
DCT on full image gives its best identification rate for only [5] S. Davis and P. Mermelstein, “Comparison of parametric representations
20*20 portion of feature vector i.e. by using only 400 DCT for monosyllabic word recognition in continuously spoken sentences,”
coefficients. DCT on image blocks gives highest identification IEEE Transaction Acoustics Speech and Signal Processing, vol. 4, pp.
rate when 16*64 portion of its feature vector is considered 375-366, 1980.
which has 1024 DCT coefficients. Finally DCT on Row Mean [6] Wang Yutai, Li Bo, Jiang Xiaoqing, Liu Feng, Wang Lihao, “Speaker
gives highest identification rate for Row Mean of 16*16 size Recognition Based on Dynamic MFCC Parameters”, International
Conference on Image Analysis and Signal Processing, pp. 406-409, 2009
image blocks i.e. for 4096 DCT coefficients. When these
highest identification rates in all three methods in DCT are [7] Azzam Sleit, Sami Serhan, and Loai Nemir, “A histogram based speaker
identification technique”, International Conference on ICADIWT, pp.
compared, it has been observed that DCT on image blocks 384-388, May 2008.
gives slightly improved results for training set of eight images [8] B. S. Atal, “Automatic Recognition of speakers from their voices”, Proc.
per speaker. Whereas, DCT on Row Mean, further improves IEEE, vol. 64, pp. 460-475, 1976.
these results with drastically reduced computations. Though the [9] Jialong He, Li Liu, and G¨unther Palm, “A discriminative training
number of coefficients used in Row Mean method is higher, algorithm for VQ-based speaker Identification”, IEEE Transactions on
overhead caused for its comparison is negligible as compared speech and audio processing, vol. 7, No. 3, pp. 353-356, May 1999.
to number of mathematical computations needed in other two [10] Debadatta Pati, S. R. Mahadeva Prasanna, “Non-Parametric Vector
approaches. Quantization of Excitation Source Information for Speaker
Recognition”, IEEE Region 10 Conference, pp. 1-4, Nov. 2008.
Similarly, WALSH on Row Mean of image gives better [11] Tridibesh Dutta and Gopal K. Basak, “Text dependent speaker
identification rates as compared to WALSH on full image and identification using similar patterns in spectrograms”, PRIP'2007
WALSH on image blocks for both the training sets. These Proceedings, Volume 1, pp. 87-92, Minsk, 2007.
better identification rates are obtained with the advantage of [12] Andrew B. Watson, “Image compression using the Discrete Cosine
Transform”, Mathematica journal, 4(1), pp. 81-88, 1994,.
reduced mathematical computations. For HAAR transform
also, identification rate for HAAR on Row Mean is better than [13] http://www.itee.uq.edu.au/~conrad/vidtimit/
HAAR on full image and HAAR on image blocks. [14] http://www2.imm.dtu.dk/~lf/elsdsr/
[15] H.B.Kekre, Sudeep D. Thepade, “Improving the Performance of Image
From the results of DCT, WALSH and HAAR on full Retrieval using Partial Coefficients of Transformed Image”,
image, it can be concluded that DCT on full image gives better International Journal of Information Retrieval (IJIR), Serials
identification rate than WALSH and HAAR on full image but Publications, Volume 2, Issue 1, pp. 72-79 (ISSN: 0974-6285), 2009.
at the expense of large number of mathematical computations. [16] H.B.Kekre, Tanuja Sarode, Sudeep D. Thepade, “DCT Applied to Row
Mean and Column Vectors in Fingerprint Identification”, In Proceedings
In WALSH transform on full image, numbers of mathematical of International Conference on Computer Networks and Security
computations required are greatly reduced as compared to DCT (ICCNS), 27-28 Sept. 2008, VIT, Pune.
since no multiplications are required in WALSH. These [17] H.B.Kekre, Sudeep D. Thepade, Archana Athawale, Anant Shah,
computations are further reduced by use of HAAR transform Prathmesh Verlekar, Suraj Shirke,“Energy Compaction and Image
but at the slight expense of identification rate. Similar Splitting for Image Retrieval using Kekre Transform over Row and
conclusions can be drawn for DCT, WALSH and HAAR on Column Feature Vectors”, International Journal of Computer Science
and Network Security (IJCSNS),Volume:10, Number 1, January 2010,
image blocks. So there is a trade off between better (ISSN: 1738-7906) Available at www.IJCSNS.org.
identification rate and less CPU time for mathematical [18] H.B.Kekre, Sudeep D. Thepade, Archana Athawale, Anant Shah,
computations. Prathmesh Verlekar, Suraj Shirke, “Performance Evaluation of Image
Retrieval using Energy Compaction and Image Tiling over DCT Row
However, in case of Row Mean approach of applying Mean and DCT Column Mean”, Springer-International Conference on
transform, performances of all three transformation techniques Contours of Computing Technology (Thinkquest-2010), Babasaheb
are same for a specific block size chosen for Row Mean. In that Gawde Institute of Technology, Mumbai, 13-14 March 2010, The paper
HAAR transform proves to be better because it requires will be uploaded on online Springerlink.
minimum number of computations. [19] H.B.Kekre, Tanuja K. Sarode, Sudeep D. Thepade, Vaishali
Suryavanshi, “Improved Texture Feature Based Image Retrieval using
The overall conclusion is that Row Mean technique Kekre’s Fast Codebook Generation Algorithm”, Springer-International
requires less number of mathematical computations and hence Conference on Contours of Computing Technology (Thinkquest-2010),
less CPU time for all three transformation techniques as

197 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
Dr. H. B. Kekre et. al.(IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 6, 2010
Babasaheb Gawde Institute of Technology, Mumbai, 13-14 March 2010, M.E./M.Tech Projects and several B.E./B.Tech Projects. His
The paper will be uploaded on online Springerlink.
areas of interest are Digital Signal processing, Image
[20] H. B. Kekre, Tanuja K. Sarode, Sudeep D. Thepade, “Image Retrieval
by Kekre’s Transform Applied on Each Row of Walsh Transformed VQ
Processing and Computer Networks. He has more than 250
Codebook”, (Invited), ACM-International Conference and Workshop on papers in National / International Conferences / Journals to his
Emerging Trends in Technology (ICWET 2010),Thakur College of credit. Recently six students working under his guidance have
Engg. And Tech., Mumbai, 26-27 Feb 2010, The paper is invited at received best paper awards. Currently he is guiding ten Ph.D.
ICWET 2010. Also will be uploaded on online ACM Portal.
students.
[21] H. B. Kekre, Tanuja Sarode, Sudeep D. Thepade, “Color-Texture
Feature based Image Retrieval using DCT applied on Kekre’s Median
Codebook”, International Journal on Imaging (IJI), Volume 2, Number Dr. Tanuja K. Sarode has received M.E. (Computer
A09, Autumn 2009,pp. 55-65. Available online at Engineering) degree from Mumbai University
www.ceser.res.in/iji.html (ISSN: 0974-0627).
in 2004 and Ph.D. from Mukesh Patel School
[22] H. B. Kekre, Ms. Tanuja K. Sarode, Sudeep D. Thepade, "Image
Retrieval using Color-Texture Features from DCT on VQ Codevectors
of Technology, Management and Engg.
obtained by Kekre’s Fast Codebook Generation", ICGST-International SVKM’s NMIMS University, Vile-Parle (W),
Journal on Graphics, Vision and Image Processing (GVIP), Volume 9, Mumbai, INDIA, in 2010. She has more than
Issue 5, pp.: 1-8, September 2009. Available online at http: 10 years of experience in teaching. Currently
//www.icgst.com /gvip /Volume9 /Issue5 /P1150921752.html.
working as Assistant Professor in Dept. of Computer
[23] H. B. Kekre, Sudeep Thepade, Akshay Maloo, “Image Retrieval using
Fractional Coefficients of Transformed Image using DCT and Walsh
Engineering at Thadomal Shahani Engineering College,
Transform”, International Journal of Engineering Science and Mumbai. She is member of International Association of
Technology, Vol.. 2, No. 4, 2010, 362-371 Engineers (IAENG) and International Association of
[24] H. B. Kekre, Sudeep Thepade, Akshay Maloo,”Performance Computer Science and Information Technology (IACSIT).
Comparison of Image Retrieval Using Fractional Coefficients of Her areas of interest are Image Processing, Signal Processing
Transformed Image Using DCT, Walsh, Haar and Kekre’s Transform”,
CSC-International Journal of Image processing (IJIP), Vol.. 4, No.2, and Computer Graphics. She has 70 papers in National
pp.:142-155, May 2010. /International Conferences/journal to her credit.
[25] H. B. Kekre, Tanuja Sarode “Two Level Vector Quantization Method
for Codebook Generation using Kekre’s Proportionate Error Algorithm” Shachi Natu has received B.E. (Computer) degree from
, CSC-International Journal of Image Processing, Vol.4, Issue 1, pp.1-
10, January-February 2010
Mumbai University with first class in 2004.
[26] H. B. Kekre, Sudeep Thepade, Akshay Maloo, “Eigenvectors of
Currently Purusing M.E. in Computer
Covariance Matrix using Row Mean and Column Mean Sequences for Engineering from University of Mumbai. She
Face Recognition”, CSC-International Journal of Biometrics and has 05 years of experience in teaching.
Bioinformatics (IJBB), Volume (4): Issue (2), pp. 42-50, May 2010. Currently working as Lecturer in department
[27] H. B. Kekre, Tanuja Sarode, Shachi Natu, Prachi Natu, “Performance of Information Technology at Thadomal
Comparison Of 2-D DCT On Full/Block Spectrogram And 1-D DCT On
Row Mean Of Spectrogram For Speaker Identification”, (Selected), Shahani Engineering College, Bandra (w), Mumbai. Her areas
CSC-International Journal of Biometrics and Bioinformatics (IJBB), of interest are Image Processing, Data Structure, Database
Volume (4): Issue (3), pp. 100-112, August 2010, Malaysia.. Management Systems and operating systems. She has 3 papers
[28] H. B. Kekre, Tanuja Sarode, Shachi Natu, Prachi Natu, “Performance in National / International Conferences /journal to her credit.
Comparison of Speaker Identification Using DCT, Walsh, Haar On Full
And Row Mean Of Spectrogram” , (Selected), International Journal of
Computer Applications, pp. 30-37, August 2010, USA. Prachi Natu has received B.E. (Electronics and
[29] H. B. Kekre, Tanuja Sarode, Shachi Natu, Prachi Natu, “Speaker Telecommunication) degree from Mumbai
Identification using 2-d DCT, Walsh and Haar on Full and Block University with first class in 2004. Currently
Spectrograms”, International Journal of Computer Science and Purusing M.E. in Computer Engineering from
Engineering, Volume 2, Issue 5, pp. 1733-1740, August 2010.
University of Mumbai. She has 04 years of
experience in teaching. Currently working as
Lecturer in Computer Engineering department
AUTHORS PROFILE
at G. V. Acharya Institute of Engineering and Technology,
Dr. H. B. Kekre has received B.E. (Hons.) in Telecomm. Shelu. Mumbai. Her areas of interest are Image Processing,
Engg. from Jabalpur University in 1958, Database Management Systems and operating systems. She
M.Tech (Industrial Electronics) from IIT has 3 papers in National / International Conferences /journal to
Bombay in 1960, M.S.Engg. (Electrical Engg.) her credit.
from University of Ottawa in 1965 and Ph.D.
(System Identification) from IIT Bombay in
1970. He has worked Over 35 years as Faculty of Electrical
Engineering and then HOD Computer Science and Engg. at
IIT Bombay. For last 13 years worked as a Professor in
Department of Computer Engg. at Thadomal Shahani
Engineering College, Mumbai. He is currently Senior
Professor working with Mukesh Patel School of Technology
Management and Engineering, SVKM’s NMIMS University,
Vile Parle (w), Mumbai, INDIA. He ha guided 17 Ph.D.s, 150

198 http://sites.google.com/site/ijcsis/
ISSN 1947-5500