19PMZD-4 Robust Speech Recognition in Both Non-Stationary Additive Noise and Reverberant

This paper addresses a problem of speech recognition in adverse noise environments. the parameter estimation for environmental compensation. In particular, the compensation of non-stationary additive noise and reverberant speech signals is a serious problem in present speech recognition techniques. To solve it, we present here a particle filter-based sequential parameter estimation me.thod for front-end environmental compensation of speech recognition in adverse environments. In the proposed method, the parameters used for environmental compensation are estimated through a sequential importance sampling step, then a residual resampling step, and finally a Markov chain Monte Carlo step with Metropolis-Hastings sampling. The estimated parameter sequence is applied to the MMSE-based clean speech signal estimation method. The evaluations were conducted on speech recognition in highly non-stationary additive noise and long reverberant time environments. In the evaluation results, we observed that the proposed method improves speech recognition accuracy over the conventional methods in both the non-stationary additive noise and rmm reverberant environments. 19PM2D-5 A Study of Talker Localization Based on Subband CSP Analysis in Real Noisy Environments