MAJOR PROJECT FINAL PRESENTATION :
TEXT PROMPTED REMOTE
SPEAKER AUTHENTICATION
Project Members:
Ganesh Tiwari (75010)
Madhav Pandey(75014)
Manoj Shrestha(75018)
Project Supervisor :
Dr. Subarna Shakya
Associate Professor
Internal Examiner:
Er. Manoj Ghimire
External Examiner
Er. Bimal Acharya
Tribhuvan University
Institute of Engineering
Pulchowk Campus
Department of Electronics and Computer Engineering
INTRODUCTION
Voice biometric system
User login
Text-Prompted system
Claimant is asked to speak a prompted(random) text
Speech and Speaker Recognition
Why Text prompted ?
Playback attack
OUR SYSTEM
Feature : MFCC
Modeling and Classifications : both statistical
GMM - Speaker Modeling :
HMM/VQ - Speech Modeling :
PROPERTIES OF SPEECH SIGNAL
Carries both Speech Content and Speaker identity
What makes Speech Signal Unique ?
Each phoneme resonates at its own fundamental frequency
and harmonics of it
Studied over short period : short time spectral analysis
What is Speaker Dependent information
Fundamental frequency, primarily
function of the dimensions and tension of the vocal chords
size and shape of the mouth, throat, nose, and teeth
Studied over long period : all the variations from that speaker
UNIQUENESS IN PHONEME
0 500 1000 1500 2000 2500-0.2
-0.15
-0.1
-0.05
0
0.05
0.1
0.15
Samples
Am
plitu
de
Phoneme /ah/
Phoneme /i:/
Pre-Processing and Feature Extraction
PREPROCESSING : STEPS
1)Silence Removal
0 1 2 3 4 5 6 7 8 9
x 104
-1
-0.5
0
0.5
1
0 0.5 1 1.5 2 2.5 3 3.5 4
x 104
-1
-0.5
0
0.5
1
Silence Signal
Silence Removed
PREPROCESSING :STEPS (CONTD..)
1)Silence Removal2)Pre-Emphasis
0 2000 4000 6000 8000 10000 120000
0.01
0.02
0.03
0.04
0.05
Frequency (Hz)
|Y(f
)|
0 2000 4000 6000 8000 10000 120000
1
2
3
4
5x 10
-3
Frequency (Hz)
|Y(f
)|
Boosted high
Frequencies
Suppressed high
Frequencies
1)Silence Removal2)Pre-Emphasis3)Framing
50% overlapped, 23ms
PREPROCESSING :STEPS (CONTD..)
1)Silence Removal2)Pre-Emphasis3)Framing 4)Windowing
0 10 20 30 40 50 60
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Hamming Window
0 200 400 600 800 1000 1200-0.04
-0.03
-0.02
-0.01
0
0.01
0.02
0.03
0.04
0 200 400 600 800 1000 1200-0.05
-0.04
-0.03
-0.02
-0.01
0
0.01
0.02
0.03
0.04
0.05
PREPROCESSING :STEPS (CONTD..)
Hamming Window
Windowed Signal
FEATURE EXTRACTION
MFCC : Mel Filter Cepstral Coefficients
Perceptual approach
Human Ear processes audio signal in Mel scale
Mel scale : linear up to 1KHz and logarithmic after
1KHz
MFCC EXTRACTION: (CONTD..)
Steps :
FFT Mel Filter Log DCT CMS
Mel Filter : 12 Filtering of absolute fft coefficients using triangular filter bank in
Mel scale
MFCC gives distribution of energy acc. to filters in Mel frequency band
Mel Filter Bank
EXTRA FEATURES :ENERGY AND DELTAS
For achieving high recognition rate
A Energy Feature
Delta and Delta-Delta
delta velocity feature
double delta acceleration feature
Co-articulation
COMPOSITION OF FEATURE VECTOR
12 MFCC Features
12 Δ MFCC
12 Δ Δ MFCC
1 Energy Feature
1 Δ Energy
1 Δ Δ Energy
39 Features from each frame
Speech Recognition/Verification by
HMM/VQ
HIDDEN MARKOV MODEL (HMM)
HMM is the extension of Markov Process
Markov Process consist of observable states
HMM has hidden states and observable symbols
per states
HMM is the stochastic model
HMM (CONTD…)
Parameters
1) The initial state distribution (π)
2) State transition probability distribution (A)
3) Observation symbol probability distribution (B)
The HMM Model
(A,B,)
EXAMPLE:
PRONUNCIATION MODEL OF WORD TOMATO
(A,B,)
HMM IMPLEMENTATION
Feature Vector observation symbols , 256
Phonemes hidden states, 6
Left to right HMM
Discrete Hidden Markov Model (DHMM) with
Vector Quantization (VQ) technique
SPEECH RECOGNITION SYSTEM
VECTOR QUANTIZATION
Speaker Recognition/Verification by
GMM
SPEAKER VERIFICATION SYSTEM
SPEAKER MODELING (GMM)
Gaussian Mixture Model
Parametric probability density function
Based on soft clustering technique
Mixture of Gaussian components
= (𝑤𝑚, 𝜇 𝑚 , 𝐶𝑚)
SPEAKER MODEL TRAINING
Estimate the model parameters
Expectation Maximization algorithm
SPEAKER VERIFICATION
Based on likelihood ratio
= 𝑙𝑖𝑘𝑒𝑙𝑖ℎ𝑜𝑜𝑑 𝑆 𝑐𝑜𝑚𝑒𝑠 𝑓𝑟𝑜𝑚 𝑡ℎ𝑒 𝑠𝑝𝑒𝑎𝑘𝑒𝑟′𝑠 𝑚𝑜𝑑𝑒𝑙
𝑙𝑖𝑘𝑒𝑙𝑖ℎ𝑜𝑜𝑑 𝑆 𝑐𝑜𝑚𝑒𝑠 𝑓𝑟𝑜𝑚 𝑖𝑚𝑝𝑜𝑠𝑡𝑒𝑟′𝑠 𝑚𝑜𝑑𝑒𝑙
TOOLS USED
Languages: Adobe Flex
Java
Blaze DS for RPC
Servers: Apache Tomcat
MySQL
Versioning Tortoise SVN
OUTPUT : SNAPSHOT (GUI)
APPLICATION AREAS
Telephone transaction
Telephone credit card purchase,
Telephone stock trading
Access control
Physical facilities
Computer networks
Information retrieval
Customers information
Forensics
Voice sample matching
LIMITATION AND FUTURE ENHANCEMENT
Noise reduction
Training on more data
Combine with
other features
other classification methods
Thanks
Any queries ?