SPEECH SILENT/UNVOICED/VOICED CLASSIFICATION WITH HIDDEN MARKOV MODELS

Open Access
Author:
Lee, Peng
Graduate Program:
Electrical Engineering
Degree:
Master of Science
Document Type:
Master Thesis
Date of Defense:
August 07, 2008
Committee Members:
  • Robert M Nickel, Thesis Advisor
Keywords:
  • silent/unvoiced/voiced speech classification
  • Fisher linear discriminant
  • Gaussian mixture model
  • hidden Markov model
  • speech processing
Abstract:
Robust silent/unvoiced/voiced (SUV) classification in noise is still considered a difficult speech processing problem. Most mechanisms proposed in the past perform well in either noise-free or with only mild additive noise environments. Their performance, however, degrades dramatically in low SNR situations. We propose an MFCC codebook based mechanism with a hidden Markov model to address the problem for a dedicated speaker scenario in a low SNR environment. Our experiments show that we can achieve a 90% correct classification accuracy in 5 dB SNR with stationary (white) noise and a 78% classification accuracy in 5 dB SNR with nonstationary (babble) noise. Our results compare favorably with the performance of a GMM based multi-features classifier and a state-of-the-art SUV classifier based on the discrete wavelet transform (DWT).