邮包校核语音识别系统的实时实现

The Implementation of a Practical Chinese Speech Recognition System For The Parcel Post Checking Task

本文研究开发了一套邮包信息校核语音识别系统 .该系统利用中大词汇量非特定人连续语音识别技术实时实现了邮包信息的语音校核 .系统可以识别普通话或四川话语音 ,可识别的词汇量约为 4 5 0 0条 .系统还采用了拒识技术与说话人自适应技术 ,提高了整个系统的稳健性 .实验表明对普通话的首选识别率达到 98.7% ,前三选识别率达到 99.9% .对四川话的首选识别率达到 95 .9% ,前三选识别率达到 98.6 % ,对无关语音的正确拒识率达到 85 % ,对口音较重的说话人经过自适应后识别率可提高 5 -

In this paper a real time Parcel Post Checking Speech Recognition System is developed.This system can achieve the parcel post information checking and verification by means of a medium vocabulary speaker independent continuous speech technology.Input speech can be Mandarin or Sichuan dialect.Vocabulary size is about 4500.By employing the advantages of the rejection model and speaker adaptation technology,the system′s robustness is improved remarkably.The experiments show that the system has a very good performance.For mandarin voice input,the recognition accuracy of the top one candidate is 98 7%,and the recognition accuracy of the top three candidates is 99 9%; For Sichuan dialect one the recognition accuracy of the top one candidate is 95 9%,and the recognition accuracy of the top three candidates is 98 6%.