Research On Speech Recognition Method In Strong Noise Environment

In recent years,speech recognition technology has been widely used in a variety of different scenarios,And the anti-noise ability has become one of them Evaluation index value of speech recognition technology system software.The voice recognition technology entity model mainly includes acoustic material entity model and language model.Obtain the function of MFCC as the input of speech recognition technology system software.DNN-HMM was the key application of the physical model of acoustic materials,and the language model was built using LSTM Internet.Under the strong noise natural environment,in the speech recognition technology system software has been built in the article.Because of the special structure of DRS model,the acoustic material solid model was created based on DRSN-HMM.DRSNHMM had good denoising performance.In the processing of sequence problems,there are many performance advantages,so that the language model was built based on that.In addition,this paper also used sequence discriminative training and speaker adaptation to optimize the acoustic model,and Word2 vec was used to generate word vectors instead of one-hot as the input of the language model.And integrated and applied NOISEX-92 noise data and thchs30 language expression data to carry out noise processing Video and voice data with different frequency stability.Finally,the characteristics of the speech recognition technology system software in a natural environment with strong noise built in the article are tested.The detection key was divided into two parts.Those of parts,Noise reduction characteristics and detect the accuracy of the speech recognition technology in the noise.It was summarized that the characteristics of different system software and analyzed the next research plan.