| Speech recognition technology has many promising applications. The application of automatic speech understanding and translation would eliminate language barrier of interaction. With the explosive extension and the rapid development of e-business, speech recognition technology will provide more convenience in many fields, including network meeting, business management, hospital, education, etc.With the rapid development of computer technology, large vocabulary speech recognition system has been implemented with high accuracy. The corpus selection has been changed from a clean environment to a noisy or reverberant environment. The research on Speech recognition under the real environment has become a hot topic.This paper analyzed the impact of the recognition accuracy of three factors: noise, reverberation and speaker overlap, based on ICSI courpus. First of all, the paper described the structure of speech recognition systems, acoustics models and language models, then analyzesd the reverberant signal and reverberant models, presents a basic method to eliminate reverberation. Then describes the structure and characteristics of ICSI Corpus .Finally, Paper proposed three algorithms to eliminate noise, reverberation and speaker overlap and carries out the experiment. When passing by spectrum subtraction and long term spectrum subtraction, the accuracy of TIDIGIT corpus recognition has improved from 64% to 91%.When passing by dominant speaker detection, the accuracy of ICSI corpus recognition rate has improved 30%.This article combines signal processing and speech recognition technology. The innovative points in combining the late reverberant variance estimation and spectrum algorithm to achieve denoised and dereverberation. Modify the ICSI corpus by dominant speaker detection to avoid overlap speaker affect on model training. |
