Research And Implementation Of Chinese Speech Keyword Recognition Algorithm

The rapid development of mobile Internet brings the greatest benefit is the convenience of information exchange,and with the convenience of information exchange has led to changes in information communication,information exchange from the initial text,graphics and so gradually to the voice,video multimedia and other forms of diversified development,therefore,voice,video and other multimedia audio key information recognition technology is particularly important.Based on the research on the development status,main technology and key problems of Chinese phonetic keyword recognition technology,in this article,the two recognition algorithms,which are supervised and unsupervised,are respectively implemented and analyzed.An improved algorithm is proposed to ensure that the recognition rate is constant and the recognition efficiency is improved effectively.This article is based on the social system module in the related scientific and technological research projects,focuses on the study and implementation of speech recognition algorithms Chinese keywords and algorithms involved in various aspects of comparative analysis and application of technology,the main work includes:(1)In this article,we researched and implemented the supervised keyword recognition algorithm based on the filler model.The semi-supervised model training method is used to make full use of the original voice without marking to improve the performance of the acoustic model.For the influence of algorithm performance,the anti-noise performance of the algorithm is improved by the combination of training the acoustic model with anti-noise performance and CMS.(2)Aiming at the shortcomings of supervised identification algorithm which can not solve the problem of less resource language recognition,in this article,we have researched and implemented the unsupervised keyword recognition algorithm based on SLN-DTW,compares the influence of the number of basic templates on the performance of algorithm recognition,The improved template fusion method reduces the time cost of the recognition process compared with the traditional score fusion method,and compares the basic data dependencies of the supervised and unsupervised algorithms.(3)Aiming at the shortcomings of the existence of time complexity and linear growth of supervisors with supervised and unsupervised recognition algorithms,based on the unsupervised identification algorithm,in this article we propose a method based on segmentation and Syllable based-DTW fast keyword recognition algorithm,by the lower dimension of the short-term spectrum characteristics of the average representation of each phoneme segment composition segment features,combined with the improved Syllable based-DTW matching method to achieve the rapid extraction of the candidate area of the keyword,the keyword candidate region based on the matching score to establish the priority queue to achieve keyword recognition.And the use of multi-process combined with the structure of shared memory to achieve the keyword recognition system automation processing.It is proved that the proposed approach can effectively reduce the recognition time,improve the recognition efficiency while ensuring the performance of keyword recognition.