Research On Speech Recognition Of Wa Language Based On Residual Network

At present,speech recognition technology develops rapidly and is widely used,but the research work on minority languages is relatively few and simple.The existing research results are mainly based on traditional phonetics.Speech recognition research is mainly based on two aspects: speech signals and spectrograms.Spectrogram is the image form of all the features of speech signal,so the recognition of spectrogram does not need to consider the voiceless voice and other features of speech itself.Residual network is a mainstream model in the field of image.This network paradigm solves the problem of network degradation.This thesis proposes a research method for wa language recognition based on residual networks,and conducts in-depth research on the related theories of convolutional neural networks and residual modules.Finally,an effective wa language spectrum classification method is obtained.The main research work and results are as follows:1)Study the theory of residual network,analyze its structural characteristics,improve the identical residual block and convolution residual block,and build a 46-layer residual network model?ResNet₄6s?.The effect of each parameter on the experimental results is verified through experiments,so as to determine the parameters of the model.Three algorithms: ResNet₅0,InceptionV3,and Inception_ResNetV2,provided by the Keras platform were selected.Through the final comparison experiments,it was found that the recognition rate of wa isolated words based on ResNet₄6s was higher than other models,and the convergence rate was faster,and the recognition rate was 96.3%.2)Develop algorithms for merging isolated words into continuous speech.Aiming at the speech corpus of wa isolated words,an algorithm?wav_merge?is proposed based on the characteristics of speech signals.In this thesis,4,000 wa continuous speech corpus were finally obtained.3)Research on wa language recognition of small-scale continuous speech spectrogram samples.Five different kinds of noise were added to each of the wa consecutive speeches.Through the final experimental results,it was found that the recognition rate of continuous speech spectrograms based on the ResNet₄6s network reached 90.2%,which proves that the model system has good robustness.