Based on the research and analysis of the construction method of traditional phonological corpus, combined with the related needs of natural spoken speech recognition and the characteristics of Tibetan natural spoken language, the construction scheme and annotation standard of spoken language corpus suitable for Tibetan speech recognition is designed. A 50-hour Tibetan Lhasa spoken corpus with five layers of annotation including phonemes, semitone, syllables, Tibetan word and sentences is also constructed. The statistic characteristics show that this corpus retains the natural properties of spoken language, and also has a balanced coverage of commonly used modeling units such as phonemes, semitone, so it is able to provide reliable data support for speech recognition technology based on Tibetan spoken speech data.%基于对普通语音语料库构建方法的研究与分析,结合自然口语语音识别研究相关需求以及藏语自然口语语音的基本特点,研究设计了适用于藏语语音识别的口语语音语料库建设方案以及相应的标注规范,并据此构建了时长50小时,包含音素、半音节、音节、藏文字以及语句共5层标注信息的藏语拉萨话口语语音语料库.统计结果显示,该语料库在保留口语语音自然属性的同时,对音素、半音节等常用语音建模单元也有均衡的覆盖,为基于藏语口语语音数据的语音识别技术研究提供了可靠的数据支撑.
展开▼
