Download Speaker Recognition System-Implementation and Applications In Computer Sciences-Project Presentation and more Slides Applications of Computer Sciences in PDF only on Docsity!
Amplitude (Normalized) Time[sec]
1
Presentation Scheme
Summary of the Previous work
Project Objectives
System Block Diagram
Research Work Techniques Used for Voice Based Person Identification and
VerificationImplemented Techniques
Project Progress
Future Direction
Amplitude (Normalized) Time[sec]
2
Project Objectives
The objectives of the project are:
Carry out Research work in the field of voice based person identification and verification
Develop a voice based person identification system that can beused for
Computer Human Interaction
Amplitude (Normalized) Time[sec]
4
Continuation from previous work…(7th^ Mid)
Project objectives and milestones
Identification taxonomy
Speaker Recognition System’s modules
Feature Extracting techniques
Mel-frequency Cepstral Coefficients (MFCC)
Feature Matching techniques
Learning Vector quantization (LVQ)
Amplitude (Normalized) Time[sec]
5
System Block Diagram
P.K Tomi, R(2004), Features Spectral Features for Automatic Text-Independent Speaker Recognition , URL: http://cis.gsu.edu/~rbaskerv/cis8680/index.html docsity.com
Amplitude (Normalized) Time[sec]
7
Speaker Recognition System’s modules
System has two modules Feature Extraction
Feature extraction is the process that extracts a small amount of data fromthe voice signal while retaining speaker discriminative information that can
Feature Matching^ later be used to represent each speaker.
Feature matching involves the actual procedure to identify the unknownspeaker by comparing extracted features from his/her voice input with the
ones from a set of known speakers.
Amplitude (Normalized) Time[sec]
8
Speaker Recognition System’s modules
Feature Extracting techniques
Mel-frequency Cepstral Coefficients (MFCC)
Linear Predictive Coding (LPC)
Linear Predictive Cepstrum Coefficient (LPCC)
Feature Matching techniques
Vector quantization (VQ)
Nearest neighbors (NN)
Gaussian Mixture Model (GMM)
Amplitude (Normalized) Time[sec]
10
MEL-FREQUENCY CEPSTRUM COEFFICIENTS (MFCC) Technique for parametrically representing the speech signalInformation carried by low-frequency components are phonetically more important for humans than carried by high frequency component. Normally 12 to 20 coefficients.The zeroth coefficient is usually dropped.
Amplitude (Normalized) Time[sec]
11
Mel-frequency Cepstrum Coefficients Processor
Computing of Mel Cepstrum
cepstrummel^ spectrum^ mel
continuous speech Blocking^ Frame frame^ Windowing^ FFT^ spectrum
Cepstrum Mel-frequency Wrapping
Amplitude (Normalized) Time[sec]
13
MFCC…
3. Fast Fourier Transform (FFT)Converts each frame of N samples from the time domain into the frequency
domain.Implement the Discrete Fourier Transform (DFT) which is defined on the set of N
samples {Xn}, as follow:
4. Mel-frequency WrappingHuman perception for frequency contents of the sound does not follow a linear
scale.Thus for each tone with an actual frequency f(Hz), a subjective pitch is measured
^ on a scale called the ‘mel’ scale.Linear spacing below 1000 Hz and logarithmic spacing above 1000 Hz.
(^1 2) /
0 ,^ 0,1, 2,...,^1
n N^ k jkn N
X k x e^ n N
(^)
mel f ( ) 2595*log 10 (1 f / 700) docsity.com
Amplitude (Normalized) Time[sec]
14
MFCC…
5. Cepstrum
In this final step, the log mel spectrum is converted back to time domain.
Inverse of DFT is replaced by taking the Discrete cosinetransform (DCT)
The result is called the mel frequency cepstrum coefficients(MFCC).
Amplitude (Normalized) Time[sec]
16
Features (^) SpeakersNo of Tested
No. ofSamples Speakerper
No. ofCorrectly MatchedSamples
PercentageResult All 19 MFCCs (1-19)
(^8) (8*4=32) 4 27 84.
Features (^) SpeakersNo of Tested
No. ofSamples Speakerper
No. ofCorrectly MatchedSamples
PercentageResult All 19MFCCs (1-19)
(^8) (8*4=32) 4 20 62.
Training results for Text-Independent SR
Testing results for Text-Independent SR
Previous Results(7th-Mid Semester)
Amplitude (Normalized) Time[sec]
17
Features (^) SpeakersNo of Tested
No. ofSamples Speakerper
No. ofCorrectly MatchedSamples
PercentageResult 10 MFCCs (^8) (8*4=32) 4 31 96.
Features (^) SpeakersNo of Tested
No. ofSamples Speakerper
No. ofCorrectly MatchedSamples
PercentageResult 10 MFCCs (^8) (8*4=32) 4 17 53.
Testing results for Text-Independent SR
Testing results for Text-Independent SR
Previous Results(7th Semester)
Amplitude (Normalized) Time[sec]
19
David Bouldin Index (DBI)
A similarity measure Rij between the clusters Ci and Cj is defined based on a measure of dispersion of a cluster Ci, Let si,
and a dissimilarity measure between two clusters dij.
max
c^1^ c c
ij^ i^ j
ij
i i n i nj ij
n c i i
R s^ s
d
R R
DB n R
^
si = Mean square distance from the points in cluster i to the center of cluster I
dij = distance between centers of clusters i and j
nc = total number of clusters
Amplitude (Normalized) Time[sec]
20
Plot of DBI for MFCCs
(^00 2 4 6 8 10 12 14 16 18 ) 10002000
30004000
50006000
70008000
9000 Plot of DBI for MFCCs
MFCCs
value of DBI
