摘要
通过窄带信道将语音信号传输到远端的识别系统,从而实现远距离的人机对话,具有重要的现实意义。在2.4kbps的速率下,语音编码算法依然可以合成出可懂度非常高的语音,但是这样的语音信号与原始语音相比还是有巨大的损失。低速率语音编码算法对于语音识别产生的影响是巨大的,因此必须想办法减轻这种损失对于识别的损害。在此选择了三种不同的低速率语音编码器,分别使用LPC(LinearPredictiveCoding,线性预测)算法、MELP(MixedExcitationLinearPrediction,混合激励线性预测)算法和IMBE(ImprovedMultibandExcitation,增强多带激励)算法,都在2.4kbps的速率下工作,将其对语音识别系统的影响进行了比较。对于特定人连续语音识别系统和非特定人连接词识别系统,在使用不同的特征矢量时,不同编码器产生的识别效果有比较大的差异。实验结果表明,语音编码器和语音识别系统的结构有很重要的联系,尽量采用相近的结构有助于获得良好的识别结果。另外,改变提取语音识别特征参数的方式也会有利于提高语音识别系统的性能。
Speech recognition through narrow band channel is important in many areas.At the bit rate as low as 2.4kbps,we can obtain synthetic speech with high intelligibility by speech coding.However,such speech signal has been damaged severely compared with the uncoded one.The low bit rate speech coding has great effect on the following speech recognition.So we must find a way to mitigate the damage of coding.We select three low bit rate vocoder working at 2.4kbps,using LPC(Linear Predictive Coding),MELP(Mixed Excitation Linear Prediction) and IMBE(Improved Multiband Excitation) algorithms respectively,to find the difference of their influences on speech recognition.Using different feature vectors,the effect of the vocoders differs on speaker dependent continuous speech recognition and speaker independent connected digits recognition.The results of the experiments have shown important relation between the algorithm structures of the vocoder and recognition system.It is helpful to use similar structures to improve the recognition performance.It will also be useful to change the way of extracting feature parameters for the recognition.
出处
《计算机应用研究》
CSCD
北大核心
2003年第9期22-25,28,共5页
Application Research of Computers
基金
国家自然科学基金资助项目(69972020)
国家基础研究项目(G1998030406)
清华大学电子工程系微波与数字通信国家重点实验室项目
关键词
语音编码
语音识别
低速率
Speech Coding
Speech Recognition
Low Bit Rate
