面向哈萨克语LVCSR的语言模型构建方法研究

一个好的语言模型不仅可以压缩语音识别过程中的搜索空间，而且还可以提高识别准确率。N-gram统计语言模型是目前广泛使用的语言模型之一。从文本的收集和处理开始，介绍了哈萨克语语言模型的构建相关技术，并以此为基础实现了一个哈萨克语连续语音识别基线系统。分别训练了基于单词和基于音节的3-gram语言模型，并通过困惑度及连续语言实验结果对两种语言模型进行了评价。

On language model construction for LVCSR in Kazakh

A good language model not only compresses the search space for speech recognition process, but also improves the recognition accuracy. N-gram statistical language model is one of the widely used language models. This paper starts from the collection and processing of the text, introduces the construction technology of Kazakh language model. On?this?basis?a Kazakh continuous speech recognition baseline system?is?implemented. It trains the 3-gram language model based on word and syllable respectively, and then the two language models are evaluated by the result of perplexity and continuous language experiment.