Dr. Anupam Shukla

1
Speaker Identification Using Wavelet Analysis and
ANN

• Dr. Anupam Shukla
• Dr. Ritu Tiwari
• Hemant Kumar Meena
• Rahul Kala

Shukla, Anupam Tiwari, Ritu Meena, Hemant Kumar
Kala, Rahul Speaker Identification using
Wavelet Analysis and Artificial Neural Networks,
proceedings of the National Symposium on
Acoustics (NSA) 2008
2
Index

1. Introduction
2. Techniques used
3. Procedure
4. Results
5. Conclusion

3
Introduction

• Identification of a person is a very traditional
  problem.
• Finger print recognition, face recognition,
  signature recognition are common techniques.
• Speaker recognition or Automatic Speaker
  Identification (ASI) identifies an author based
  on the words spoken.
• We have used wavelet analysis to extract the
  various features and Artificial Neural Networks
  to identify the speaker by the extracted features.

4
Common Techniques

• Analysis techniques
• Fourier Analysis
• Short Time Fourier Analysis
• Wavelet Analysis
• Artificial Neural Networks

5
Analysis Techniques

• We have used Wavelet transform to extract
  characteristics, which is an advancement over
  Fourier analysis and Short Time Fourier Analysis
  (STFT).

6
Fourier Analysis
7
Short Time Fourier Analysis
8
Wavelet Analysis

• It is a windowing technique with variable-sized
  regions.
• Wavelet analysis allows the use of different time
  intervals for different type frequency
  information.

9
Wavelet Analysis(Cont..)
10
Wavelet Analysis(Cont..)

• Capable of revealing aspects of data
• Wavelet packet method
• Signal decomposition

11
Wavelet Packet Analysis
12
Artificial Neural Network

• Excellent means of machine learning
• Reputed training of the system to learn the given
  data
• Testing
• Performance

13
Procedure

• Collection of data sets
• Analysis of data sets (feature extraction)
• Training of ANN
• Testing
• Result

14
Normalization of Data

• Ii(Vi - Mean(Vij) ) / (Max(Vij) - Mean(Vij) ),
  for all j
• Here
• Ii is th ith input of the neural network
• Vi is the ith feature extracted from Wavelet
  Analysis
• Mean(Vi) is the mean of all Vij found in the
  training data set
• Max(Vi) is the maximum of all Vij found in
  training data set for all j in data set

15
Feature Extracted
16
Result

• Performance of 97.5
• This clearly shows that the algorithm works well
  and gives correct results on almost all inputs.
• 20 speakers and 40 test cases (39 correctly
  identified)

17
Conclusions