论文部分内容阅读
采用基于段长分布的非齐次隐马尔可夫模型(DDBHMM)进行维吾尔语声学建模。在新语料下由于总词汇量的成倍增加导致识别时间倍增,为缩短识别时间将耗时最长的概率计算部分采用多线程机制优化了识别模块,同时加入了端点检测进行控制,并相继设计了录音模块、特征提取模块、波形显示及结果输出显示模块等,对这些模块进行集成界面化后产生了一个基于DDBHMM的维吾尔语连续语音声学层实时识别系统,并对系统进行了测试及验证.
Uyghur-language acoustic modeling is performed using the non-homogeneous Hidden Markov Model (DDBHMM) based on the segment length distribution. Under the new corpus, the recognition time is doubled due to the doubling of the total vocabulary. In order to shorten the recognition time, the computation of probability takes the longest. The multi-threading mechanism is used to optimize the recognition module. At the same time, endpoint detection is added to control the recognition, A recording module, a feature extraction module, a waveform display and a result output display module are integrated. After these modules are integrated and interfaced, a real-time recognition system of Uyghur continuous speech-acoustic layer based on DDBHMM is generated, and the system is tested and verified.