以维吾尔语为例研究自然语料缺乏的民族语言连续语音识别方法。采用HTK通过人工标注的少量语料生成种子模型,引导大语音数据构建声学模型,利用palmkit工具生成统计语言模型,以Julius工具实现连续语音识别。实验用64个维语母语者自由发话的6 400个短句语音建立单音素声学模型,由100 MB文本、6万词词典生成基于词类的3-gram语言模型,测试结果表明,该方法的识别率为72.5%,比单用HTK提高4.2个百分点。 A Case Study of Uyghur Language for Continuous Speech Recognition in Ethnic Languages. Using HTK, a few corpus of human annotation were used to generate the seed model, guide the large voice data to construct the acoustic model, use the palmkit tool to generate the statistical language model, and use the Julius tool to achieve continuous speech recognition. Experiments were conducted on 6 400 short sentences phonetic of 64 speakers of free speech. A 100-gram text and a 60,000-word dictionary were used to generate a 3-gram language model based on parts of speech. The test results show that the recognition method The rate was 72.5%, 4.2 percentage points higher than HTK alone.