基于动态时轴弯曲的汉语连呼数字识别算法

A Dynamic Time-Warping Algorithm for Chinese Continuous Digit Recognition

模板匹配是计算机语音识别的重要手段之一。本文以孤立数字的线性预测编码(LPC)参数为模板,针对当汉语连呼数字发音速度较快时一字分割较困难这一特点,研究了不需预分割的汉语连呼数字识别问题。算法通过把复合距离的端点搜索与端点无约束的动态时轴弯曲(DTW)技术相结合,在保证一定识别率的前提下,使计算量较二级或多级动态规划算法有了大幅度下降即接近孤立数字发音识别水平。同时,简化的树状决策方法使识别性能进一步得到提高。初步实验结果表明,此方案可较有效地克服连呼数字识别中的连读效应。

Template matching is one the important technology in the area of computer speech recognition. In this paper,an improved non-segmenting approach to Chinese connected digit recognition with a set of LPC-based, isolated word reference patterns is presented to counter the diffculty of segmentation in case of fasts speaking rate. Compared with Two-level DP an level building DTW,the computational complexity is largely reduced by the combination of the endpoint defection of compound distance with unconstrainted dynamic time-warping (DTW) under a prerequisite of no drop in the recognition accuracy. A sim-plied tree-shaped decision is used to improve the performance of the system. The preliminary results show that the method is effective in the coarticulation of connected digit recognition.