摘要
为了在未知一段语音所属语言种类的情况下将其转换为正确的字符序列,将语种辨识(language identification,LID)同语音识别集成在一起建立了中、英文大词汇量连续语音识别(large vocabulary continuous speech recognition,LVCSR)系统。为了在中、英文连续语音识别系统中能够尽早的对语音所属的语言种类做出判决以便进行识别,从而降低解码的计算量,对语种辨识过程中的语种剪枝进行了研究,表明采用合理的语种剪枝门限在不降低系统性能的情况下,可以有效的降低系统的计算量及识别时间。
In order to transfer the speech into the correspond text without knowing the language, the language identification （LID） is integrated into speech recognition and then the large vocabulary continuous speech recognition （LVCSR） system is developed which support English and mandarin. The language pruning during the LID is discussed for making decision which language the sp6ech belong to early, then the speech can be recognized and the calculation is reduced in decoding. The experiments show that, if the pruning threshold is set reasonable, it could decrease the calculation, and so the system output the recognition result more quickly without losing the performance.
出处
《计算机工程与设计》
CSCD
北大核心
2007年第8期1931-1933,共3页
Computer Engineering and Design
基金
国家863高技术研究发展计划基金项目(2001AA114071)
关键词
连续语音识别
语种辨识
段长分布
非齐次隐含马尔科夫模型
语种剪枝
continuous speech recognition
language identification
duration distribution
inhomogeneous hidden Markov model
language pruning
