基于HMM／VQ的认人的中等词表连续语音识别

本文讨论基于隐马尔可夫模型(HMM)和矢量量化(VQ)的连续语音识别方法。用这种方法,对每个单词作成一个HMM,对多个模型组合成的状态转移网络搜索其状态转移的最佳路径,从而实现不预先进行单词切分的连续语音的识别,使用有限态文法约束及其它一些改善识别性能的措施,演示系统能识别特定人的18种英语句式,150个单词,用312个话句(共有2710个单词)进行测试,识别延迟时间为发音时长的62％,发音速度平均为每秒2.32个单词,单词识准率为97.3％。

Speaker-dependent Medium Vocabulary Continuous Speech Recognition Based on HMM and VQ

In this paper a method of continuous speech recognition based on hidden Markov models (HMM) and vector quantization (VQ) is discussed. According to the method, forming its own HMM for each word's voice, searching the optimal path of state transition network combined by HMM' s of all words of the vocabulary, it has been realized to recog-nize continuous spoken sentences without presegmentation of each word. The recognition performance is improved by using finite-state syntactic analysis and other techniques. The de-monstration system with vocabulary of 150 words can process 18 type of English sentences. The delay from speech end to getting recognition result is nearly 0.62 times as long as the duration of speech. Tested by 312 sentences including 2710 words, the word recognition accuracy is 97.3% when the average speech speed is 2.32 words per second.