EEL 6586-Automatic Speech Processing Hidden Markov Models for Speech Recognition

1
EEL 6586-Automatic Speech ProcessingHidden
Markov Models for Speech Recognition

• Savyasachi Singh
• Computational NeuroEngineering Lab
• March 28, 2007

2
Introduction
3
Model Parameters
4
Assumptions
5
Three basic problems
6
Evaluation Problem
7
Forward Algorithm
8
Backward Algorithm
9
Decoding Problem
10
Viterbi Algorithm
11
Learning Problem
12
ML Estimation EM algorithm
13
Baum Welch Algorithm
14
Re-estimation formulae
15
Gradient based method
16
Practical Pitfalls
17
Limitations
18
Isolated Word Recognition
HMM Word 1
HMM Word 2
FEATURE EXTRACTION
SELECT MAXIMUM
HMM Word 3
HMM Word K
19
Typical Implementations
20
HW 4 part c pseudocode

• Chop speech signal into frames and extract
  features. (preferably MFCC)
• Choose HMM parameters N, M, cov. type, A etc.
• Start learning procedure for train set
• for each word repeat following steps
• for each state
• Initialize GMMs and get parameters (use
  mixgauss_init.m)
• end
• Train HMM with EM (use mhmm_em.m)
• end
• Start testing procedure for test set
• for each test utterance
• Compare with all trained models and get log
  likelihood (score) using forward backward
  algorithm. (use mhmm_logprob.m)
• Select model with highest score as recognized
  word.
• end
• 5. Tabulate confusion matrix.