基于贝叶斯方法的鲁棒语音切分

Bayesian Approach to Robust Speech Segmentation

在基于隐马尔科夫模型的语音切分基础上 ,融合了不受噪声干扰的先验切分模型 ,提出了基于贝叶斯方法的语音切分方法。在贝叶斯切分方法的框架内 ,作者首先对语音序列进行了变换 ,将由切分点构成的序列变为由音节长度构成的序列。然后 ,假设音节长度序列符合一阶马尔科夫过程 ,经过归一化处理后 ,求出了切分的先验概率公式 ,得到了贝叶斯方法的切分模型。在噪声环境下的实验证明 ,由于切分模型独立于噪声 ,对在噪声环境下声学模型的失配提供了很好的补偿 ,使得语音切分的鲁棒性大大增加。

A large quantities of phonetically segmented speech is necessary in speech recognition and speech synthesis. Based on HMM speech segmentation, this paper presents a Bayesian approach to robust speech segmentation. By hidden Markov model, it combined with a prior segmentation model which is independent to noise feature as the compensation for the mismatch of acoustic model to enhance the robust performance. Experiment shows the validity of Bayesian speech segmentation.