利用声调判别提高汉语数码语音识别性能

Use tone detection to improve performance of mandarin digit speech recognition

数码语音“２”与“８”等的混淆是汉语数码语音识别错误的主要来源。为此，提出了在汉语数码语音识别中引入声调判别的方法。首先设计了变长度短时平均幅度差函数（ＬＶＡＭＤＦ）、元音中心定位、基音谐波单周期校正以及基音邻近搜索等一系列高性能基音周期估计算法，在此基础上设计了一个针对汉语数码语音声调识别的ＭＤＴＤ算法。实验表明，新的基音周期估计方法和ＭＤＴＤ算法使汉语数码语音识别率由９５．２％上升到９８．５％，更使“２”与“８”的分辨率由９０．５％上升到了９８．８％，从而较好地解决了这个难题。

Confusion between digits such as “2” and “8” has been the main error source in mandarin digit speech recognition (MDSR). Tone detection is introduced into MDSR to solve this problem. A series of methodologies for high performance pitch contour estimation are developed, including length varied average magnitude difference function (LVAMDF), vowel center location, multi period to single period pitch adjustment, pitch neighborhood searching, etc. The mandarin digit tone detection (MDTD) algorithm is then designed for MDSR tone detection. Experiments show that the new methodologies and algorithms increase MDSR correct recognition rate from 95.2% to 98.5% , and improve the correct recognition rate between digit “2” and “8” from 90.5% to 98.8%, thus basically remove this confusion from MDSR.