线性预测分析在连接词语音识别中的研究

Research of Linear Prediction Analysis in Conjunction Speech Recognition

特征参数的提取是关系到语音识别系统性能好坏的关键,而线性预测分析是目前普遍采用的特征参数提取方法。针对在连接词和连续语音识别系统中,传统的线性预测系数已不能满足特征提取的要求,研究采用了三种主要的线性预测推演参数,即线性预测反射系数、线谱对系数和线性预测倒谱系数,及其在连接词语音识别系统中的应用,并进行计算机仿真。仿真结果表明,在输入语音库与信噪比一致的情况下,线性预测倒谱系数的识别率最高。从而证明,在包含语义特征信息和说话人特征方面,线性预测倒谱系数性能要优于线谱对系数和线性预测反射系数。

Extraction of feature parameters is related to the performance of speech recognition system,and the linear prediction analysis is currently widely used feature extraction methods.In conjunction and continuous speech recognition system,the traditional linear prediction coefficient（LPC） has been unable to meet the requirements of feature extraction.This article discusses three main parameters of the linear predictive inference,they are linear prediction reflection coefficient（LPRC）,line spectral pairs（LSP） and linear prediction cepstrum coefficient（LPCC）,and the applications in conjunction speech recognition system.Through computer simulation,its performances are analyzed and compared.The simulation results show that when the ratio of signal to noise is accordant with the input speech corpus,the linear prediction cepstrum coefficient is of the highest recognition rate.Therefore,the semantic features,such as contains information and speaker characteristics,linear prediction cepstrum coefficient,are superior to the performance of line spectral pairs and reflection coefficients.