一种新的全汉语单音节语音识别算法

A New Method for Chinese Sylables Speech Recognition

本文阐述了一种新的全汉语单音节语音识别算法一DP/MVQ法.新方法充分借鉴了隐马尔可夫模型(Hidden Markov Model)中“状态”的概念,保留了David K.Bordon提出的多段矢量量化(Multisection Vector Quantization,简记为MVQ)方法中能保持时间序列信息的优点,并且在码本的训练过程中用了动态规划(Dynamic Programming)技术去优化MVQ产生的码字,使得DP技术贯穿于码本训练和识别的全过程。新方法充分考虑了汉语普通话语音的声学结构特点和统计特性,而且训练和识别均较快,码本尺寸也较小.新方法着重于基于DSP硬件的实时实现,以便于能用语声控制汉字的计算机输入。

This paper illustrates a new algorithm - -DP/MVQfor Chinese sylables recognition task. The method cites from the concept of 'STATE' in Hidden Marcov Models (HMM) , it reserves the advantage of preserving time sequence information in multisection vector quantization (MVQ) method proposed by David K. Bordon, and Dynamic Programming technique is used during codebook training , this makes DP technology used from training to recognizing . The method pays much attention to the structure and statistical character of ordinary Chinese speech. It not only decrease the size of pattern database , but also makes recognition more faster . The method contributes to real-time recognition based on DSP hardware , so that we can input Chinese words freely by speaking.