基于音节的维吾尔语大词汇连续语音识别系统

Syllable based language model for large vocabulary continuous speech recognition of Uyghur

维吾尔语是一种黏着语,基于单词的语言模型不太适合于维吾尔语大词汇连续语音识别任务。该文提出了适合维吾尔语的基于音节的语言模型,引入最大匹配分词算法评价音节语言模型在大词汇连续语音识别任务中的单词识别性能。实验结果表明:基于音节的语言模型在未登录词和模型复杂度等方面表现出比基于单词的语言模型更加优越的性能,并且使识别系统的单元错误率比基于单词的系统减少了50%。因此,在维吾尔语语音识别任务上可以将音节作为识别单元。

Uyghur is an agglutinative language so word based language models are not the best tools for Uyghur large vocabulary continuous speech recognition（LVCSR） systems.This paper presents a syllable based language model,which is better suited to Uyghur,using a maximum matching word segmentation algorithm to evaluate word recognition performance of the syllable based language model in the LVCSR task.Tests show that the syllable based language model outperforms word based language models in terms of the number of out of vocabulary words（OOVs） and model complexity and that the syllable based system reduces the relative unit error rate by 50% over the word based system.Therefore,Uyghur syllables can be used as recognition units for some speech recognition tasks.