论文部分内容阅读
本文研讨缺乏语言资源的民族语言(如维吾尔语)中如何引用语音技术、开发应用系统问题.提出基于GMM-UBM混合SVM技术方法实现实用性说话人识别系统,通过小语料人工标注语音语料预选高精度声学根(seed)模型、再引导大语料训练生成鲁棒性声模提高连续语音识别精度实现汉民会话语音翻译系统.对维吾尔语70人发话电话语音识别实验结果显示,基于GMM–UBM–SVM方法的不特定说话人识别实验其正确识别率为94.3%,比先行GMM–UBM方法精度提升3%;基于seed声模HTK-Julius技术的维吾尔语连续语音识别实验,其识别率为72.5%,比直接使用语音文本对齐语料单靠HTK实现识别方法(63.2%)精度提高9.3%;同时本研究讨论基于Moses技术的汉维医院门诊会话语音翻译系统预测Blue值达到了57.7%.
This paper discusses how to apply the speech technology and develop the application system in the minority languages (such as Uyghur language) which lack of language resources.This paper proposes a practical speaker recognition system based on GMM-UBM hybrid SVM technology, Precision acoustics seed model, and then guide the large corpus training to generate a robust acoustic model to improve the accuracy of continuous speech recognition to achieve a Chinese-speaking voice translation system.According to GMM-UBM-SVM The accuracy of this method is 94.3%, which is 3% higher than that of prior GMM-UBM method. The recognition rate of Uyghur continuous speech recognition based on seed acoustic model HTK-Julius is 72.5% Compared with the direct use of speech text alignment corpus HTK recognition method (63.2%) accuracy increased by 9.3%; at the same time, this study discusses the Moses technology outpatient session speech translation system predicts Blue value reached 57.7%.