基于HMM的声调语音模型研究

针对声韵母相同但声调不同的近音字识别问题和声韵母及声调都相同的同音字识别问题，提出在语音模型和语言模型中分别引入声调和字转移概率，以提高近音字和同音字的识别率。首先将声调划分为5种表现形式添加到汉语音节的最后一个音素中构成新音素，使用高斯混合隐马尔科夫模型建模新音素。然后通过统计方法计算特定语境下的字间转移概率。最后使用HTK工具包实现了带声调的语音模型和有字转移概率的语言模型。实验结果证明添加声调可以提高近音字的识别率，使用特定语境下字间转移概率可以提高同音字的识别率。

HMM-Based Tone Speech Model

To improve the recognition rate of approximant characters with the same initial but different tones and the recognition accuracy of the homophonous characters with the same initial and tone， we introduced the tone and word transition probabilities into the models of speech and language respectively. Firstly， the tone is divided into five forms and added to the last phoneme of Chinese syllable to form a new phoneme， which was afterwards modeled by Gaussian mixed hidden Markov model. Then， we calculated the word transition probabilities in a specific context. Finally， we adopted the Hidden Markov Model Toolkit to realize the models of tonal speech and language with word transition probabilities. The experiments show that the tones can improve the recognition rate of approximant characters， and the use of word transition probabilities in a specific context can promote the recognition rate of homophonous characters.