基于特征参数归一化的鲁棒语音识别方法综述

目前,自动语音识别系统往往会因为环境中复杂因素的影响,造成训练环境和测试环境存在不匹配现象,使得识别系统性能大幅度下降,极大地限制了语音识别技术的应用范围。近年来,很多鲁棒语音识别技术成功地被提出,这些技术的目标都是相同的,主要是提高系统的鲁棒性,进而提高识别率。其中,基于特征的归一化技术简单而有效,常常被作为鲁棒语音识别的首选方法,它主要是通过对特征向量的统计属性、累积密度函数或功率谱的归一化来补偿环境不匹配产生的影响。该文主要对目前主流的归一化方法进行介绍,其中包括倒谱矩归一化方法、直方图均衡化方法以及调频谱归一化方法等。

Survey of Feature Normalization Techniques for Robust Speech Recognition

The performance of current automatic speech recognition (ASR) systems often deteriorates radically when the input speech is corrupted by various kinds of noise sources. Such performance degradation is mainly caused by mismatch between the training and recognition environments. Quite a few techniques have been proposed to reduce this mismatch over the past several years. Some of the techniques, like feature-based normalization, are generally simple yet powerful to provide robustness against several forms of signal degradation. So normalization strategies are often chosen as the preferred method for speech robustness. They are employed by normalizing the statistical properties (moment), cumulative density function or power spectral density (PSD) of feature vector to compensate for the effects of environmental mismatch. In this paper, most commonly used feature normalization methods are reviewed, such as cepstral moment normalization, histogram equalization technique (HEQ) and Modulation Spectrum Normalization etc.
Key wordsrobust speech recognition; cepstral mean normalization; high order cepstral moment normalization; histogram equalization; cepstral shape normalization