多语言语音识别声学建模理论和容错识别新方法研究

The multilingual speech recognition is one of the key unresolved technologies in the field of speech recognition. This application mainly focuses on the innovative research for the speech recognition acoustic modeling theory for the speech data resource-constrained small languages, and the multilingual error-tolerant speech recognition decoding method. Main contents are following: (1) Investigate the articulatory feature extraction approaches, articulatory features are more general than the phonemes, to improve and increase the data reusable capability. (2) Research on the structural feature extraction and the discriminative transformation methods based on the multilayer neural network, in order to obtain more robust and discriminative features. (3) Investigate the acoustic modeling methods based on the shared multilingual model sets and the subspace modeling for low-resource languages, in order to realize the robust acoustic modeling of multilingual and low-resource language speech recognition. (4) Investigate the data-credibility evaluation method, data using rules, and feedback-based discriminative learning theory, in order to use the more sophisticated model discriminative learning. (5) Research the multilingual speech recognition method based on the embedded language identification, and an error-tolerant speech recognition algorithm, in order to develop a multilingual speech error-tolerant recognition decoding engine. (6) Finally build a multilingual speaker-independent continuous prototypal speech recognition system, then evaluating and improving a multilingual speech recognition system and a speech recognition system of low-resource language under the realistic conditions. This research is of important theory significance and practical application values.

多语言语音识别是语音识别领域中尚未解决的关键技术之一。本申请主要针对语音数据资源受限的条件下的小语种语音识别建模理论和多语言容错语音识别方法展开创新性研究。主要的研究内容包括：（1）研究比音素更一般化发音属性特征提取方法，提高语音数据复用能力。（2）研究基于神经网络细胞器的结构化特征提取方法和鉴别性特征变换方法，提高特征的鉴别性和稳健性。（3）研究基于多语言模型集共享的声学建模和基于子空间小语种语音声学建模方法，实现多语言和小语种语音识别声学模型的稳健建模。（4）研究音特征数据可用度估值方法和数据使用规则和基于反馈的鉴别性学习理论，实现精细的模型鉴别性学习。（5）研究基于内嵌语种识别的多语言语音识别解码方法和容错识别算法，实现多语言语音容错识别。（6）最后构建多语言非特定人连续语音识别原型系统，在真实条件下，对多语言和小语种语音识别系统性能进行测试和改进。该研究具有重要理论意义与实用价值。

低资源和多语言语音识别是信息处理领域的桂冠，它涉及认知科学、声学、信号处理、机器学习等众多学科，推动多语言语音识别的进步可以促进相关学科的深入交叉与融合。.本项目面向语音识别领域中多语言语音识别，针对语音数据资源受限的条件下的小语种语音识别建模理论和多语言容错语音识别方法展开创新性研究。课题执行过程中，课题组研究了比音素更一般化发音属性特征提取方法，提高了语音数据复用能力；研究了基于神经网络细胞器的结构化特征提取方法和鉴别性特征变换方法，提高了特征的鉴别性和稳健性；研究了基于多语言模型集共享的声学建模和基于子空间小语种语音声学建模方法，实现了多语言和小语种语音识别声学模型的稳健建模；研究了语音特征数据可用度估值方法和数据使用规则和基于反馈的鉴别性学习理论，实现了精细的模型鉴别性学习；研究了基于内嵌语种识别的多语言语音识别解码方法和容错识别算法，实现了多语言语音容错识别；构建了多语言非特定人连续语音识别原型系统，参加NIST关键词检测（OpenKWS）学术评测，并于2015年获得公开条件国际第一名。.本项目共发表论文70篇，其中SCI检索11篇，授权发明专利3项。培养博士后1名，博士生8名（其中2名在读），硕士生4名。