Research And Implementation Of Noise Classification Based Speech Enhancement

In speech signal processing system,noise signal can cause serious degradation of the system performance.Speech enhancement is used to filter the noise to improve the quality of the speech.However,in the real-world environment,there are many types of noise.The diversity of noise characteristics greatly limits the performance of speech enhancement algorithm,and reduces the quality,especially the intelligibility of enhanced speech.Therefore,a speech enhancement system based on noise classification is proposed in this paper.The proposed consists of three parts: Voice Activity Detection(VAD),noise classification and noise power spectrum estimation.Firstly,VAD is performed to recognize the noise frames and the speech frames.A VAD algorithm based on Long-term Power Spectrum Variability(LPSV)is proposed in this paper.In this algorithm,the power spectrum variation of signals over a long period of time is used as the feature to distinguish noise frames from speech frames.The proposed VAD algorithm can achieve higher accuracy in different noise and SNR conditions.Especially in the non-stationary noise environment,the accuracy of voice activity detection is significantly improved.Then,the noise classification model is used to identify the noise type.In this paper,the noise classification model based on Convolution Neural Network is proposed.In this model,Mel-Frequency Cepstral Coefficients of multi-frame noise signals are used as features to classify types.In the non-interference environment,the average classification accuracy of the proposed noise classification model is 98%,in the interference environment,the average classification accuracy is 85%.Finally,according to the results of noise classification,the optimal combination of parameters is selected to estimate the noise power spectrum using the Improved Minima Controlled Recursive Average(IMCRA)algorithm,and Optimally ModifiedLog-Spectral Amplitude(OM-LSA)is used for speech estimation.Compared with the traditional speech enhancement algorithm based on IMCRA,the proposed algorithm can improve the quality of the noisy signal,especially the intelligibility of the noisy signal.The performance of the proposed algorithm was tested on TIMIT corpus and Noisex-92 noise database.From the spectrogram,it can be seen that the enhanced speech signal retains more details of the speech signal while de-noising.The objective metrics include Short-Time Objective Intelligibility(STOI),Perceptual Evaluation of Speech Quality(PESQ)and Segmental Signal-Noise Ratio(Seg-SNR).Compared with the traditional IMCRA algorithm,the objective evaluation metrics of enhanced speech signal is improved obviously.