用共振峰轨迹提高汉语数码语音识别性能

Use formant trajectory to improve the performance of mandarin digit speech recognition

在汉语数码语音识别（ Ｍ Ｄ Ｓ Ｒ）中，“２”和“８”是最易混淆的一对语音。文章分析了“２”和“８”混淆的原因，发现可用于分辨“２”和“８”的区别特征在于其共振峰轨迹的差异。因此文章提出了基于共振峰轨迹的判决算法（ Ｆ Ｔ Ｂ Ｄ）来分辨“２”和“８”。实验表明，使用 Ｆ Ｔ Ｂ Ｄ 算法，使 Ｍ Ｄ Ｓ Ｒ识别率从９６０％ 提高到 ９７７％ ，“２”和“８”的识别率从 ９１％ 提高到９９％ ，消除了这对语音的混淆，提高了 Ｍ Ｄ Ｓ Ｒ

In mandarin digit speech recognition (MDSR), “2” and “8” are the most confusable pair of words. The reason why “2” and “8” are often confused is analyzed. It is found that the cue to distinguish “2” and “8” is the difference between the formant trajectory of “2” and “8”. Therefore the formant trajectory based on decision algorithm (FTBD) was proposed to distinguish “2” and “8”. Experiments show that with FTBD the correct recognition rate is improved from 96.0% to 97.7% for MDSR, and from 91% to 99% for “2” and “8”, thus this confusion is removed from MDSR, and the performance of MDSR is improved.