ENHANCED POWER-NORMALIZED FEATURES FOR MANDARIN ROBUST SPEECH RECOGNITION BASED ON A VOICED-UNVOICED-SILENCE DECISION

Power-normalized features have been shown to improve the performance of English large vocabulary continuous speech recognition under different acoustic conditions. In this paper, considering tone characteristics of Mandarin speech, we adopt different strategies to deal with different sounds based on a voiced-unvoiced-silence decision of sounds. For voiced sounds, harmonic enhancement based on a weighted harmonic-noise-model (WHNM) is applied to accurately capture the salient harmonic information and decreases the effect of various non-stationary noises. After this, standard power-normalized processing (SPNP) is performed. For unvoiced sounds, the SPNP is only used. For silence sounds, an quality frame dropping (FD) algorithm is incorporated into the front-end properly. As a result, enhanced powernormalized features are obtained and used to process noisecorrupted Mandarin speech. The experimental results show better recognition accuracies for Mandarin continuous speech recognition in noisy environments over the ETSI Advanced Front-End (AFE).