Research On Continuous Speech Recognition Technology In Noisy Environment

Artificial intelligence(AI)is one of the most popular research fields now,and speech recognition is an important research direction of AI.With the development of science and technology,continuous speech recognition technology has made great progress.At present,the recognition rate of continuous speech recognition system has reached a high level in the laboratory environment,but in noise environment,the recognition rate will be greatly affected.In the practical application of continuous speech recognition system,noise is almost unavoidable.Therefore,the research of continuous speech recognition technology in noise environment is particularly important.Based on the theoretical basis of speech recognition technology,this paper introduces the components of continuous speech recognition system,including speech signal preprocessing,speech signal feature analysis,continuous speech segmentation,acoustic model and language model.Finally,a medium vocabulary Chinese continuous speech recognition system in noise environment is realized,and the performance of the system is tested.The main contents of this paper are as follows:(1)Analysis of speech signal feature.Firstly,this paper studies the speech signal preprocessing technology,including speech enhancement technology based on spectral subtraction.Then,this paper analyses the characteristics of Chinese speech signal in time domain,frequency domain and cepstrum domain,and extracts a variety of feature parameters.Among them,the frequency domain-based spectrogram and the Cepstrum-based pitch track are emphatically studied.(2)Continuous speech segmentation technology.Segmentation of continuous speech includes two steps,one is endpoint detection,the other is segmentation of speech segment primitives.Firstly,this paper studies the multi-threshold endpoint detection technology based on time-domain feature parameters.Then,based on the analysis of pitch track and spectrogram,a Chinese continuous speech syllable segmentation algorithm with certain anti-noise is studied.Experiments show that the continuous speech segmentation algorithm has high accuracy.(3)Analysis of speech recognition model.Continuous speech recognition system is divided into two layers,acoustic layer and language layer.This paper studied three acoustic recognition models,which are Vector Quantization Model,discrete Hidden Markov Model and continuous Hidden Markov Model.The recognition rates of the three models are compared using experimental speech database.Then the influence of different training samples on the recognition rate of continuous Hidden Markov Model is studied.Then,the N-gram model is studied as the recognition model of language layer.A complete continuous speech recognition system is established by combining these two layers.