采用模糊切分实现非特定人连接数字语音识别

SPEAKER INDEPENDENT CONNECTED-DIGITS SPEECHRECOGMITION USING FUZZY SEGMENTATION

选用倒谱特征、△倒谱特征、能量类特征等多个语音信号特征，采用多码本和ＨＭＭ／ＶＱ技术实现非特定人连接数字语音识别。在识别阶段，使用帧同步型网络搜索识别算法结合模糊切分实现非特定人连接数字语音识别。数字串识别准确率达到８１％以上，而数字识别准确率更高，达到９４．４％。

In this peper we describe a speaker-independent connected-digits Chinese speechrecognition system in which the well-known technique of vector quantization (VQ) and hidden Narkovmodeling(HMM)are combined with multi-features including cepetra, delta cepetra and energies. Anapproach of fuzzy transition is included in the frame-spochronous network search recosnitionprocedure to handle transition between digit medels. This system achieved a string accuracy of81. 8%,and a digit accuracy of 94. 4%.