论文部分内容阅读
在一个基于经典隐马尔可夫模型（Hidden Markov Model,HMM）的汉语全音节、非特定人、连续语音识别系统中,利用声学层分数和基于拼音的统计语言模型分数,对关键词的可信度进行贝叶斯估计。本文提出了最大后验（Maximum APosteriori,MAP）可信测度,给出了计算 MAP可信度分数的前向后向算法。并且在关键词捕捉应用中评价了 MAP可信测度的性能,实验表明MAP可信度分数对关键词候选具有很强的鉴别能力。此外,MAP可信测度可以广泛地应用于各种语音识别应用中。
In a Chinese syllable, non-specific and continuous speech recognition system based on Hidden Markov Model (HMM), using the acoustics layer and Pinyin-based statistical language model scores, Bayesian estimation. In this paper, we propose a Maximum A Posteriori (MAP) credible measure and give a forward-backward algorithm for calculating the MAP confidence score. And the performance of MAP credible measure is evaluated in the keyword capture application. The experiment shows that the MAP credibility score has a strong ability of discriminating the keyword candidates. In addition, MAP Trusted Measures can be used in a wide range of speech recognition applications.