基于多词汇树的对话语音识别搜索策略

在查询类对话语音识别系统中，词汇中的少数几类词往往占有词汇量的很大部分，利
用这一重要特点，可以对传统的基于单树结构的搜索算法进行改进。本文在采用基于类的语言模型的条件下，在查询类对话语音识别系统中，对词汇中的若干大类单独建树，从而免除了多数词汇的语言模型look-ahead 计算时间和存储空间，并实现了语言分的提前精确运用。针对机票预订口语对话系统的实验表明，该搜索策略与单树方法相比，在同样的计算资源条件下，时间效率有显著提高。
关键词: 语音识别 对话系统 搜索 词树 look-ahead

Abstract A few of classes often occupy large proportion of vocabulary in speech recognition for many information query dialog system, which can be used to improve traditional search algorithm based on single lexicon tree. Under the condition of using class based n-gram language model, this paper presents a new search strategy which builds separate trees for some big word classes ( classes that own many words) to dispenses with look-ahead computation of many words in decoding process and to use precise probabilities of language model earlier. The experiments on flight reservation dialog system show that the time efficiency of the presented method is obviously higher than single-tree method under equal computational cost.
Key Words speech recognition; dialog system; search; lexicon tree; look-ahead