摘要:语音识别技术是一门交叉学科，正逐步成为信息技术中人机交互的关键技术。通过语音传递信息是人类最重要、最有效、最常用和最方便的交换信息形式。本文是对特定人连续电话号码的语音进行识别。本系统中采用基于短时平均幅度和过零率的双门限法进行语音端点检测；用充分考虑了人耳听觉特性的Mel频率倒谱参数(MFCC)作为语音信号特征矢量，采用隐马尔可夫模型(HMM)算法来完成语音模板的训练和语音识别的任务。对电话号码语音识别系统进行验证，实验室环境现场识别100组随机电话号码，其中全部识别正确的有22组。通过对0~9这10个数字分别统计，最低识别率为72.84%。
关键词：语音识别；端点检测；MFCC；HMM
Abstract:The speech recognition is a cross-discipline, and is gradually becoming the key-technology of human-machine interation in information technology. Transmission of information through the voice of humanity’s most important, most effective,most popular and most convenient form of exchange of information. This paper presents the speaker-dependent continuous speech recognition of telephone numbers. This system is based on SAM-SAZR(Short-time Average Magnitude and Short-time Average Zero-crossing Rate). The system adopts Mel Frequency Cepstrum Coefficient (MFCC), which considers fully sense of hearing characteristic, as voice signal eigenvector. The discrete hidden markov model(HMM)is adopted to train and recognize the speech signal. Experiments are done to validate the telephone voice recognition simulation, on-site to identify 100 groups of random numbers, which identify all right with 22 groups. Through the 10 Numbers 0~9 were statistics, minimum recognition rate is 72.84%.
Key words：Voice Signal, Endpoint detection, MFCC,HMM
本课题基于Matlab仿真软件实现特定人电话号码连续语音识别系统的设计。首先，利用Windows自带的录音设备进行录音，完成了对模型库的训练，建立了0~9这个10个数字的模型建立。其次，通过对语音识别的基本原理和模式识别基本原理的学习，更好的理解了语音识别的过程。本文采用了MFCC特征提取和HMM识别算法，完成了对11位和12位电话号码的识别。因为手机号码是11位，而固定电话的位数是区号+8位号码，例如常州的区号为0519，而南京的区号025。当捕捉到的电话号码不足11或者超过12位时，系统会给出选择是否继续录音。
