Speaker Identification CS5984 Pattern Recognition and Classification Course Project

1
Speaker IdentificationCS5984 Pattern Recognition
and Classification Course Project

• Seonho Kim, Seungwon Yang

2
Outline

• Problem Definition
• Data Set
• Resources Used
• Seven Steps of Our Approach
• HMM Parameters
• Demo
• Conclusions

3
Speaker ID

• Problem definition
• One of the forms of biometric identification
• Method to recognize a person based on voice
• Two Types
• Text dependent
• Text independent

4
Data Set

• 5 utterances of Virginia Tech from 5 persons
  recorded in quiet place
• 4 utterances for training HMM
• 1 utterance for evaluation

5
Resources Used

• HMM Toolbox from http//www.cs.ubc.ca/murphyk/Sof
  tware/HMM/hmm.html
• MFCC module from VOICEBOX http//www.ee.ic.ac.uk/h
  p/staff/dmb/voicebox/voicebox.html
• Signal Preprocessing Tools from MIT
  http//web.mit.edu/sharat/www/resources.html

6
Data Collection

• STEP1 Data Collection
• Recorded a word 5 times per each speaker.
• 5 speaker.

7
Preprocessing

• STEP 2 Silence Removal
• Chop out silence parts

8
Preprocessing

• STEP 3 Emphasize signal
• Amplify the signal

9
Feature Extraction

• STEP 4
• Mel Frequency Cepstral Coefficient (MFCC)
• 12 features extracted

10
Feature Extraction

• STEP 5
• Shift signal to from 1 to 100

11
HMM Training

• STEP 6 HMM Training Implemented Four different
  training versions
• 12 features Using MFCC
• 1 selected feature One selected coefficient
• MFCC Mean Average of 12 coefficient
• Iterative speaking Virginia Tech X 4 (MFCC
  mean)

12
Evaluation

• STEP 7 Compute the log-likelihood of testing
  data using trained HMMs
• Choose HMM with largest log-likelihood
• Identify the speaker of the HMM

13
HMM Parameters

• Number of visible states
• Discretized signal (Observed number of integers
  between 1 to 100)
• Number of hidden states
• Arbitrary number (15)
• Sample size
• First 30 samples (in iterative speaking model)
• Whole single utterance (in other 3 models)

14

• Demo!

15
Conclusions and Future Works

• Our HMM based Speaker Identification System Shows
  average 6070 of accuracy in problems of picking
  one speaker out of five
• Four different Training Methods show similar
  performance
• We will further develop this system to be Text
  Independent

16

• Questions / Comments?