Effectiveness analysis of anti-aliasing feature extraction for performance improvement of automatic speech recognition

Abstract

A variety of compensation techniques have been proposed for robust speech recognition. However, conventional feature extraction methods and feature enhancement techniques do not consider aliasing noise in the feature extraction. The aliasing noise in the feature domain can occur according to insufficient frame rate. The feature extraction process using a frame shifting scheme can be viewed as a sampling process in discrete-time domain. In the signal sampling, an anti-aliasing process should be carried out based on the Nyquist sampling theorem to avoid aliasing noise. Therefore, speech features also should be extracted with the anti-aliasing process. In this thesis we analyze the effectiveness of the anti-aliasing feature extraction method that reduces the aliasing noise in the feature domain by extracting the feature sequence with small frame shift, low-pass filtering the feature sequence, and down-sampling at the frame rate same as that in the conventional feature extraction. In addition, we deal with how to determine filter types and parameters used in the anti-aliasing feature extraction. Furthermore, we investigate the combination of the anti-aliasing feature extraction and cepstral mean normalization (CMN) methods. Because two methods are independent on each other in their effects, we can expect substantial improvement in performance. Various evaluations were conducted to verify the effectiveness of the anti-aliasing feature extraction method. Experimental results show that anti-aliasing feature extraction method is very effective for large phone context task. Also, the results show that IIR type low-pass filter (LPF) is more suitable for the anti-aliasing filter in speech recognition systems. Finally, we examined the effectiveness of the anti-aliasing feature extraction in noise environments. 5.6 - 13.6 % of error reduction rate (ERR) was observed in additive noise environments and 16.9 - 19.8 % of ERR was observed in convolutive noise environments. In c...