Method of speaker adaptive speech recognition

Abstract

The patent describes a noise-tolerant method of spectral analysis processing suitable for use in a speech recognizer system. The method, referred to as minimum mean log spectral distance (MMLSD), proceeds by estimating a frame of clean speech for each noisy speech frame. This is done using mixture model hidden Markov models of the clean speech signal and using the assumptions that different noise frequency channels are mutually uncorrelated; and that different speech frequency channels are uncorrelated within each state of the Markov process.mDLR 43.72.Ne TRAINING MODULE FOR ESTIMATING MIXTURE GAUSSIAN DENSITIES FOR SPEECH-UNIT MODELS IN SPEECH RECOGNITION SYSTEMS Yunxin Zhao, assignor to Matsushita Electric Industrial Company