无端点检测汉语识别算法的实现及改进——动态时间规整和隐马尔可夫统一模型的应用

A recognition algorithm without the ending point detection of chinese based on the DTW and HMM unified model

语音识别算法中，动态时间规整（ＤＴＷ）和隐马尔可夫模型（ＨＭＭ）是最有效的识别算法，并且两者之间有着本质的联系和内在的统一［１］，据此前期工作中，已经建立了ＤＴＷ和ＨＭＭ的统一模型（ＤＨＵＭ）［２、３］。本文对ＤＨＵＭ进行了改进，在ＤＨＵＭ中引进寂静段自环，并根据汉语语音的特点，提出了一种无端点检测的语音识别算法。在识别过程中，该算法无需确定语音信号起止点位置，而是从寂静段开始，直接按帧提取特征（帧长２０ｍｓ，帧间重叠５０％），特征向量由１５阶倒谱系数和帧平均能量组成。实验中，用ＤＨＵＭ实现了该算法，对９９个相似汉语单字的识别测试结果表明：无端点检测的识别正识率为９４．９５％，正识率下降很少，但不作端点检测却降低了算法的复杂程度。为进一步改善识别性能，特征向量采用一种听觉模型特征，识别器具有更好的鲁棒性，识别率会略有提高。

In speech recognition, dynamic time warping(DTW) and hidden Markov model(HMM)are the most effective algorithm,and there are intrinsical relations between them . According to this, an unified model of DTW&HMM has been established in previous work. In this paper, by introducing the self loop of the stationary segment of the DTW and HMM Unified Model (DHUM) ,and according to the characteristic of Chinese speech, a recognition algorithm without the ending point detection is proposed. Compared with the traditional method, in this algorithm, there is no necessary to decide the ending point of speech signals. From the stationary segment on, feature vectors, which consist of 15 order cepstrum coefficients and the average energy of each frame, are extracted in frames(length of each frame is 20 millisecond, the overlapping between two frames is 50%), this algorithm is successfully implemented. In recognition of 99 similar words of Chinese, a first candidate recognition rate of 94 95% is obtained. If an auditory feature is accepted for feature vectors, the robustness of the algorithm will be better.