摘要
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%.
In this paper, we report our recent researches in speech techniques, such as speaker identification （SI）, continuous speech recognition （CSR） and speech translation （ST）, among the minority languages spoken in China, like Uyghur, Khasak and Mongolian. We tried an approach GMM-UBM-SVM for real calling SI, and HTK-Julius for CSR by Uigur and Mongolian, and Moses software for ST for medical service. Experimental results show that accuracies of 94.3%for SI with GMM-UBM-SVM, and of 72.5%for CSR by Uygur, and showed a Blue value of 57.7%for the medical service ST by Uyghur using Moses software.
出处
《新疆大学学报（自然科学版）》
CAS
2014年第1期88-96,共9页
Journal of Xinjiang University(Natural Science Edition)
基金
国家自然科学基金(61163030)
自治区科技支疆项目(201291116)资助
关键词
语音技术
民语
说话人识别
连续语音识别
语音翻译
GMM-UBM-SVM
Speech technology
speaker identification
continuous speech recognition
speech translation
GMM-UBM-SVM
