一种人机交互语音切分系统

对于大规模的语音语料,语音切分方法主要有传统的人工切分和机器自动化切分2种方式.人工切分大规模语音语料的切分质量易控制,但效率低、成本高;机器自动化切分效率高,但后期查找切分错误时任务极其繁重.因而提出一种人机交互语音切分系统,切分人员可选择自动切分算法,设置切分参数,修改有问题的自动切分结果,同时可自动生成用于HTK训练的标注文件.以课题组采集的1 000个普米语语音文件为研究对象,以普米语孤立词为切分基元,机器自动化切分存在难以避免的切分错误,后期检查时工作量巨大;然而使用本文提出的人机交互语言切分系统进行切分,切分人员在无需高认知度的情况下也可做到近100%的切分正确率.

A system of speech segmentation based on human-computer interaction

The methods of speech segmentation are divided mainly into the manual segmentation and the machine-automatic segmentation for large speech corpora. The quality of segmentation can be controlled easily through the traditional manual segmentation; however,the shortcomings of manual segmentation such as inefficiency,and high cost are also obvious. As we all know,the method of machine- automatic segmentation has the advantage of high efficiency,but there is much work in detecting the segmentation errors. Thus,this paper proposes a system of speech segmentation based on human- computer interaction which provides different segment algorithms and parameters,modifies the errors of automatic segmentation results,and generates the labeling of HTK training files. Taking one file that includes one thousand units of the Pumi speeches as the research object,this research has proved that a personal with low segmentation- related cognitive competence can use the proposed segmentation system to achieve nearly one hundred percent accuracy.