摘要
本文给出了一个高性能汉语数码串非特定人连续语音识别系统 ,其声学模型基于Mel倒谱系数和连续HMM ,识别时采用多候选帧同步搜索算法 ,并采用了MCE算法进行训练以提高系统的区分能力 ,实验证明该系统的识别率为 94 8% (不定长数字串 )和 96 8% (定长数字串 ) .为增强系统的实用性 ,本文还研究了基于MAP算法的说话人自适应算法和基于置信度的拒识算法 .在进行自适应后 ,误识率可相对下降 40 %以上 ,在拒绝掉 5 %的正确语音时 ,系统识别率可以上升到 96 9% (不定长数字串 )和 98 7% (定长数字串 ) .
A high performance mandarin digit string speaker-independent speech recognition system is given. The acoustic model is based on the Mel Frequency Cepstrum Coefficient and the continuous hidden Markov model (HMM). The multi-candidate frame synchronous search algorithm is adopted in the recognition stage with the MCE algorithm as the training approach. Experiments demonstrate that the correct recognition rate of the system is 94.8% (unknown length) and 96.8% (known length). In order to enhance the flexibility of the system, this paper also conduct research on the maximum a posteriori (MAP) based speaker adaptation and confidence measure based rejection. More than 40% recognition errors can be removed after adaptation and the recognition rate can be improved to 96.9% (unknown length) and 98.7% (known length) when 5% of the correct results are rejected.
出处
《电子学报》
EI
CAS
CSCD
北大核心
2001年第5期595-599,共5页
Acta Electronica Sinica
基金
国家自然科学基金! (No .699750 0 7)
国家 863项目! (No .863 30 6ZD1 3 0 4 6)
关键词
汉语数码串
语音识别
语音信号处理
Acoustic signal processing
Algorithms
Mathematical models
