基于分段模型的连续语音识别

近年来在大词汇连续语音识别的研究取得了长足的进步，隐马尔柯夫模型（HMM）是连续语音识别的核心部分。但是HMM对语音信号的描述不完善，为此人们提出了很多替代模型，其中一类将语音信号描述为长度随机的特征矢量序列，称为随机分段模型（Stochastic Segment Models)，简称为分段模型（SM）。该文将首先阐述分段模型的原理，并将分段模型和隐马尔柯夫模型进行比较，其次给出基于分段模型的识别和模型训练算法，最后给出实验结果并进行了讨论。

Speech Recognition Based on Stochastic Segment Models

Recently-peat progress has been made in the field of Large Vocabulary Continuous Speech Recognition, and HMM stands the kernel. However, HMM has its own shortcoming when describing the speech signals. Many alternative modelshave been proposed and one approach named Stochastic Segment Models (SSM) represented the speech signal with a variable - length sequence of observation vec- tors. Here we describe the details of the segment Model and give the recognition and taming algorithms, and discuss some practical result in the final part.