Research On Chinese Syllable Evaluation Approach After Automatic Speech Recogniton

Nowadays,automatic speech recognition (ASR) technology develops quite quickly and is being used in many domains.However, there still are so many problems.Because of background noises,accents of different persons and some new words,the precision of ASR in real environment is not desired.Moreover, speech document retrieval has wide application now. There are plenty of texts after ASR need to be processed on a large scale. Therefore, we use post-processing for texts after ASR to improve ASR effects.This paper is mainly based on science and technology research key projects of Ministry of Education,"The key technology research in robust understanding of results after speech recognition",which aims to analyze the results of speech recognition to make them more accurate and complete.In this background, the thesis focuses on the completion of a whole speech-recognition text processing system and post-processing information of text evaluation method for ASR research.Main contents of this thesis and my work are:1.The creation of Chinese syllable knowledge base.I made phonetic N-gram model and the introduction and comparative analysis of data smoothing of more appropriate model for the Chinese syllable and data smoothing method.I also described how to use different syntax to build models and smoothing method of phonetic knowledge base.2.The positioning of anchor words in speech recognition results.I made continuous phonetic N-gram model analysis and evaluation, sum up the corresponding laws, auxiliary applications syllable stability and end up with the determination of anchor words.3.Suggestions for correction candidate words.I describe phonetic confusion rules in the corresponding phonetic similarity and convert them into the corresponding candidate words.In order to carry out scoring the candidate words by phonetic similarity to measure the extent of the closest and the original spelling.4.Application of algorithms to a real system.The above algorithm is applied to of to a real whole error detection and correction after speech recognition systems.We introduce the framework and concrete realization of their systems to explain role of this work and a comparative analysis of the practical application of results.