论文部分内容阅读
研究汉语自然口语识别中的建模单元选择问题。在HMM三状态模型中,声韵母单元与音素单元作为两种最流行的建模单元各有优劣。一方面从自然口语音变严重的问题出发,倾向采用粗粒度的声韵母单元以概括各种音变;另一方面从三状态结构可能无法有效描述复杂单元的问题出发,又倾向采用细粒度的音素单元。本文在实验语音学理论研究成果与声韵母时长分析实验结果的基础上,主张对扩展声韵母单元进行有选择的拆分,提出了基于鼻韵尾分离的声韵母拆分方法。实验结果表明本文的方法与扩展声韵母单元、音素单元相比,识别性能有了明显改善,其字错误率分别降低2.23%和9.45%。
Research on the Selection of Modeling Units in Chinese Natural Spoken Chinese Recognition. In the HMM three-state model, vowels and phonemes are the two most popular modeling units each having advantages and disadvantages. On the one hand, starting from the problem of serious natural speech changes, we tend to use coarse-grained vowel units to summarize various sound changes. On the other hand, starting from the problem that the three-state structure may not be able to describe complex units effectively, Phoneme unit Based on the experimental phonetics theory and vowel duration analysis, this paper advocates the selective disassembly of extended vowel units and proposes a new separation method of vowels based on nose-end separation. The experimental results show that the proposed method has a better recognition performance than the extended vowel unit and phoneme unit, with the word error rate reduced by 2.23% and 9.45% respectively.