An Extended Study on Signal Bias Removal For Telephone Speech Recognition

Abstract

Speech signals transmitted over telephone channel often suffer from interferences of additive ambient noise and multiplicative distortion. In this paper, a proportional alignment decoding (PAD) instead of Viterbi decoding (VD) is proposed and combined with a modified signal bias removal (MSBR) method to improve the performance and computation efficiency of a telephone speech recognition system. This proposed method simplifies the computation in decoding process and uses only the features in cepstral domain. Experiments on multispeaker ( 50 males and 50 females) isolated Mandarin digit recognition are conducted to demonstrate the effectiveness of the proposed method. Experimental results shoe that a remarkable improvement in recognition accuracy up to 20%-30% at 0 dB SNR can be achieved. Moreover, due to its simplicity in state alignment, the PAD method requires much less computation time in state decoding, and is more flexible in real-time applications.