Monte Carlo Noisy HMM Estimation and Segmental Differential Features on the Aurora2 Clean Training Evaluation

In this paper, the compensation of mismatch between clean training hidden Markov models (HMMs) and noisy test speech is addressed. The purpose is to approach the performance of Aurora2 multi-condition training but use only clean training material. The idea is to integrate three methods including (1) mean subtraction, variance normalization and ARMA filtering (MVA) post-processing for Mel-scaled cepstral coefficients (MFCCs) normalization, (2) Monte Carlo noisy HMM estimation by adding artificial noises in the linear mel-scale filterbank parameter (MELSPEC) domain and (3) novel segmental differential features for increasing recognizer’s discriminative power. Experimental results on Aurora2 clean training corpus have shown that great performance improvement was achieved. Especially, although only clean training material was used, the performance did close to the level of Aurora2 multi-condition training.