MMSE estimation of speech power spectral density under speech presence uncertainty for automatic speech recognition

Abstract

In order to improve the performance of automatic speech recognition (ASR) system in noisy environment, this paper first derives a minimum mean square error (MMSE) speech power spectral density (PSD) estimator. Besides, speech presence uncertainty (SPU) is explored to obtain further improvement. The speech spectral amplitude model assumed in this paper is Chi distribution instead of the traditional Rayleih distribution. Simulation experiments of ASR system demonstrate that the new proposed approach outperforms the traditional estimation of speech spectral amplitude and PSD with SPU under Rayleih distribution.