一种引入帧间相关信息的HMM语音识别方法

A METHOD OF HMM SPEECH RECOGNITION INTRODUCED INTER-FRAME CORRELATION

该文提出了一种基于复数帧段输入HMM的语音识别方法,它采用相继的复数帧组成的特征参数向量作为语音识别HMM的输入,能有效地在语音识别HMM中引入帧间相关信息。为了进一步改善复数帧段输入HMM的输出概率分布函数,作者还提出了用MGDF和RBF函数作为复数帧段输入HMM的输出概率分布函数的方法。通过对非特定人汉语孤立数字和连续数字语音识别试验,证实了该文提出的引入帧间相关信息方法的有效性。

This paper applies segmental unit into HMM for speech recognition. In this model, several successive frames are combined and treated as an input vector. It expects that segmental unit input HMM would be effective to describe the inter-frame correlation information and has also proposed the MGDF and RBF to further improve output probability function. By comparing them with the traditional HMMs based on their speech recognition performance rates through the experiments of speaker-independent spoken digit (isolated/connected) recognition, the validity of the proposed appraoch could be verified.