Savyasachi Singh Computational NeuroEngineering Lab March 19, 2008
Introduction
Model ParametersModel Parameters
Assumptions
Three basic problemsThree basic problems
Evaluation ProblemEvaluation Problem
Forward AlgorithmForward Algorithm
Backward AlgorithmBackward Algorithm
Decoding ProblemDecoding Problem
Viterbi AlgorithmViterbi Algorithm
Learning ProblemLearning Problem
ML Estimation: EM algorithmML Estimation: EM algorithm
Baum Welch AlgorithmBaum Welch Algorithm
Re-estimation formulaeRe-estimation formulae
Gradient based methodGradient based method
Practical PitfallsPractical Pitfalls
Limitations
Isolated Word RecognitionIsolated Word Recognition FEATURE EXTRACTION FEATURE EXTRACTION HMM Word 1 HMM Word 1 HMM Word 2 HMM Word 2 HMM Word 3 HMM Word 3 HMM Word K HMM Word K SELECT MAXIMUM SELECT MAXIMUM
Typical ImplementationsTypical Implementations
HW 4 part c pseudocodeHW 4 part c pseudocode 1.Chop speech signal into frames and extract features. (preferably MFCC) 2.Choose HMM parameters N, M, cov. type, A etc. 3.Start learning procedure for train set for each word repeat following steps for each state Initialize GMM’s and get parameters (use mixgauss_init.m) end Train HMM with EM (use mhmm_em.m) end 4.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.