On robust feature representation for speech recognition in adverse environments

Abstract

The problem of noise robustness is one of the most important issues for comercializing of speech recognition systems. This dissertation details the development of robust feature representation for the speech recognition in adverse environments. The basic aim is to remove slow-varying noise and speaker-specific components by filtering of feature parameter sequence. While conventional high-pass approaches use a band-pass or a high-pass filter in the feature parameter domain, the proposed methods introduce the decorrelation principle to suppress noise components and to satisfy the observation independent assumption of hidden Markov model (HMM). This decorrelation principle is implemented as a temporal filter to provide an alternative of conventional filtering methods. First, according to the decorrelation principle, a novel filter design method for high-pass approaches was proposed. This decorrelation technique derived a well structured high-pass filter, and the Wiener filtering was added to suppress the artifacts introduced by a overlapped frame analysis. Thus, the resulting filter was implemented as a band-pass filter, which attenuates low modulation frequencies. The proposed frame decorrelation processing (FDP) effectively de-emphasized noise components, and confirmed the effect of high-pass approaches with a theoretical justification. In order to perform the FDP, the power spectrum of the feature sequence was first estimated, and the error bounds due to a feature analysis were extracted. Then, the FDP provided a band-pass filter using the obtained power spectrum and error bounds. The experimental results indicated that the FDP outperformed other methods for a noisy speech recognition. Note that sufficient states for each HMM are required. Since high-pass approaches attenuate the stationary regions, this may be critical in the stationary-based recognizer. Compared to the delta feature with only transitional information, the FDP included both instantaneous and ...