论文部分内容阅读
在连续语音中,不同的说话者在不同语境下说话的速度差异是很大的。偏离正常语速往往会造成识别错误,使识别性能下降。考虑到语速对于语音单元段长的影响是同步增长或同步下降的,相邻语音单元的段长之间存在很强的相关性,本文从利用段长的相关信息出发,在基于段长分布的隐含马尔可夫模型(DDBHMM:Duration Distribution Based HMM)的框架上,提出了一种语速自适应算法。对数字串和大词汇量连续语音识别的试验表明这个算法是有效的。
In continuous speech, the difference in speed at which different speakers speak in different contexts is significant. Deviations from normal speech rate often result in recognition errors, degrading recognition performance. Considering that the effect of speech rate on speech segment length is synchronous or synchronous, there is a strong correlation between segment lengths of adjacent speech elements. Based on the relevant information of segment length, (DDBHMM: Duration Distribution Based HMM) framework, a speech adaptation algorithm is proposed. Experiments on digital string and large vocabulary continuous speech recognition show that this algorithm is effective.