自适应帧长特征提取对语音识别系统性能的影响

The Influence of Adaptive Frame-length Feature to Speech Recognition Performance

提出一种自适应帧长语音特征分析方法，使语音编码更准确，达到提高语音识别性能的目的。该方法包括过渡帧检测和过渡语音帧特征表示两方面。采用了两种特征表示方法。基于ＴＩＭＩＴ语音数据包和自定义的汉语语音数据的单词识别实验表明，这两种表示方法有相同的效果，都能在一定程度上提高识别系统的性能，但计算量稍有区别。基于ＴＩＭＩＴ数据的ＤＨＭＭ系统和ＣＨＭＭ系统的错识率分别下降了１１．２１％和９．５８％；基于自定义数据的ＤＨＭＭ系统和ＣＨＭＭ系统的错识率分别下降了１１．５５％９．５％。

in this paper, a speech analysis approach with adaptive frame length is proposed to solve the shortcoming of fixed frame length speech analysis, which could not provide optimal coding for every events of speech. Transient signal is detected upon spectral and temporal characteristics of speech. Two expression schedules are used to represent the feature of a transient frame. Word recognition experiments on both TIMIT and NTIMIT databases showed that the proposed speech analysis could significantly improve recognition performance, but the extla computation cost is very little. On TIMIT database, word classification with DHMM (CHMM) demonstrated a 11.21% and 9.58% error rate reduction compared with fixed frame length. and comparable results are achieved for NTIMIT database.