多码本HMM汉语语音识别

MULTI-CODEBOOK HMM METHOD FOR CHINESE SPEECH RECOGNITION

本文用多码本结构的HMM进行汉语语音识别。语音特征有Cep(倒频谱)系数及其一阶差分ΔCep。建立特征矢量Cep1、Cep2和ΔCep1的三个VQ码本,分别训练出三个HMM模型。实验结果表明此方法优于以LPC为特征参数的单码本HMM方法。由于增强了辅音的特征,对易于混淆的孤立字语音取得90％的识别率。对二字词组语音的识别率为98％。

This paper presents an approach of the multi-codebook HMM for Chinese speech recognition. The cepstrum and its first order differential are two parameters for the recognition system. Three codebooks are coeated for the vectors Cep1, cep2 and△cepl, then three HMM models are trained. Experimental results show this method is better than the HMM based on a single codebook using L P C. The characteristics of the consonants have been enhanced, this leads to the 90% recognition rate for the isolated words easy to be confused. The 98% rate also is oftained for two words composition of Chinese speech.