普通话连续数字串语音识别的持续时间模型

Duration Modeling for Continuous Mandarin Digital Speech Recognition

在普通话连续数字串的识别中 ,与传统 HMM在持续时间模型上的错误假设有关的删除与插入错误所占比例可达 2 4 .2 3% .基于此 ,在 Viterbi解码中引入持续时间模型信息 .对多种带参函数分布的持续时间模型在理论和实验上的比较分析都证明了 Gamma分布更能精确反映汉语字模型的持续时间特性 .文中还在外惩罚模型的基础上提出了预加重分段内惩罚持续时间模型和全局内惩罚持续时间模型两种改进算法 .实验表明 ,结合持续时间模型的语音识别算法可以有效地减少删除与插入错误率 ,使总体识别错误率比基带系统减少了 47.74% .

In a continuous Mandarin digit recognizer,the insertion and deletion errors related to the conventional HMM's false assumption on duration modeling amount to 24.23% in all recognition errors.This paper applied duration information into Viterbi decoding to overcome these errors. All the theoretic analysis on different parametric distributions and experiment results conclude that Gamma distribution comes out optimally characterize syllable level duration in Mandarin. In addition to ex penalty function, two forms of durational model were proposed: pre weighted in penalty function and global penalty function. The experimental results indicate that combining durational model with traditional recognition algorithm can effectively reduce both the deletion and insertion error rate and consequently about 47.74% total recognition error rate reduction is achieved over the baseline system.