Robust HMM phoneme modeling for different speaking styles

Abstract

The authors describe the robustness of six types of phoneme-based HMMs (hidden Markov models) against speaking-style variations. The six types of models are VQ (vector quantization)-based and fuzzy VQ-based discrete HMMs, and single-Gaussian and mixture-Gaussian HMMs with either diagonal or full covariance matrices. The mixture-Gaussian HMM with diagonal covariance matrices, the fuzzy VQ-based discrete HMM, and the single-Gaussian HMM with full covariance matrices show better results than the other three in 18-Japanese-consonant recognition experiments. The authors also propose a model-adaptation technique that combines multiple models using the deleted interpolation. This technique makes models easy to apply to different-speaking-style speech. >