Research On Novel Robust Speech Recognition

Noise robustness is one of the major obstacles to the commercial use of speech recognition techniques. Most of the current speech recognizers are designed to work in controlled environments using clean speech. Though they can achieve rather high recognition rates in these environments, their performance degrades rapidly when noises exist. This phenomenon is mainly caused by the mismatches between training and testing conditions and makes the current speech recognizers unsuitable for many real world applications. In this dissertation, techniques for robust speech recognition are discussed. The main achievements are listed as follows:1. Conventional spectral compression scheme would over-compress some frequency components and under-compress other frequency components of speech signal at the same time. To solve this problem, we propose a Perceptual Non-uniform Spectral Compress(PNSC) technique according to the power law of hearing. After carrying out analyses and experiments, there is a points worthy to mention that PNSC scheme is more robust than the conventional feature extraction scheme. Substantial improvement over the conventional feature extraction scheme is found under low SNR noise conditions.2. Inspired by the partial masking by background noise on a sound signal, a SNR-dependent Non-uniform Spectral Compression (SNSC) method is proposed to better simulate the intensity-to-loudness conversion. Recognition results show that the SNSC front-end can deal with different types of additive noises with performance significantly better than that of the standard LPCC, MFCC and PLPCC front-end and other NSC scheme. The analysis for model parameter from different front-end demonstrates that the SNSC derived feature is more robust than the conventional feature. Moreover the choose for parameters in SNSC is analyzed and given in the thesis.3. When applying the SNSC scheme for a real recognition task, the speech models need to be retrained. To solve this drawback, we present a novel Model-base Adaptation based on SNR Non-uniform Spectral Compression(MA-SNSC). Based on some assumptions, the thesis gives the procedure of deriving the algorithm. Experimental outcome demonstrates that the SNSC scheme is a robust speech...