基于深度迁移学习的安多藏语语音识别研究

安多藏语是藏语的三大主体方言之一，从语音识别角度看，属于低资源或者极低资源语言。关于安多藏语语音识别，仅有个别学者进行了探索性研究，研究内容分散且不够深入。因此，项目组拟开展基于深度迁移学习的安多藏语语音识别研究。本项目主要进行以下四个方面的研究：第一，进一步完善已有的安多藏语语音样本数据库；第二，从藏语为低资源语言的现状出发，研究基于深度迁移学习的安多藏语语音特征提取；第三，构建基于注意机制、对抗机制的深度神经网络声学模型；第四，构建基于Transformer的语言模型。本项目主要利用深度迁移学习、注意机制、对抗机制等最新模式识别技术，重点解决低资源条件下面向安多藏语的特征提取、声学模型构建、语言模型构建等关键科学问题，构建高效、准确的安多藏语语音识别系统。积极开展本项目研究有利于丰富低资源语言语音识别研究，实现藏语的语音识别输入，完善藏语的人机交互，保证网络中传播的藏语信息内容的安全。

Tibetan is a low resource language and Amdo Tibetan is one of its three main dialects. Up to the present, only a few researchers have done some exploratory researches on Amdo dialect speech recognition, and there is no systematical, effective, and practical Tibetan speech recognition system. Therefore, this research team will devote itself to this challenging project. Firstly, the team will construct a speech database of Tibetan Amdo dialect. Secondly, the team will explore some deep transfer learning-based feature extraction methods for low-resource language such as Amdo Tibetan. Thirdly, the team will construct an effective acoustic model based on attention mechanism and adversarial mechanism. Fourthly, the team will construct a Transformer-based language model for Tibetan..This research team will mainly focus on feature extraction, acoustic modelling, and language modelling for low resource language such as Amdo Tibetan by enrolling such novel techniques as deep transfer learning, attention mechanism, adversarial mechanism, etc. In this way, the team could set up a systematic, effective, and practical speech recognition system for Amdo Tibetan..The study of this project have many great theoretical and social significances. It can benefit the theory development of low resource language speech recognition, can perfect human-machine interface related to Tibetan language, and can ensure the security of Tibetan language information in internet. Furthermore, this project can train a group of talented researchers for Tibetan speech recognition.