New Speech Features Based on time-varying LPC for Robust Automatic Speech Recognition

摘要

Discrimination of similar pronunciation phrases is more difficult than that of normal phrases. In this paper, we propose the use of fast Fourier transform (FFT) based mel frequency cepstral coefficients (MFCCs) with time-varying linear predictive coding (TVLPC) based cepstrum. Due to the presence of intra-frame variations in TVLPC, it is expected that this approach will improve speech recognition. Evaluation results demonstrate that the proposed approach achieves between 46.67% and 60% and between 76.67% and 86.67% using dynamic range adjustment (DRA) as noise suppression technique at 10 dB and 20 dB SNR respectively. This is in comparison to 43.33% and 70% recognition accuracy achieved using conventional approach under similar conditions.