Speech Enhancement Algorithms For Suppressing Musical Noise In Spectral Subtraction

In real life,the transmission and recording of speech signals will inevitably be affected by additive noise and reverberation,which is not conducive to the processing of speech signals and the extraction of features,and also reduces the clarity and intelligibility of speech signals.Suppressing noise and reverberation by the speech enhancement algorithms to improve the quality of the speech signal as much as possible to restore the original clean speech signal from the speech signals with noise and reverberation.Among the speech enhancement algorithms,spectral subtraction has become one of the most widely used algorithms due to its good denoising and dereverberation effect,fewer constraints and simple operation.However,the musical noise generated in the spectral subtraction reduces the effect of the algorithm and affects the quality of speech.Target to spectral subtraction,this thesis presents some algorithms to suppress the musical noise:(1)The masking effect is used to suppress the musical noise generated by the spectral subtraction dereverberation.According to the fact that musical noise is generated by the spectral subtraction algorithm,the difference in the frequency domain for the enhanced speech signal with musical noise and reverberation speech signal is performed,weighting coefficients are combinined to obtain the masking factor,then the masking factor is used to process the enhance speech signal to reduce the musical noise.In the simulation,the de-reverberation signal and the output result are compared by using the time domain waveform,the spectrogram and the perceptual evaluation of speech quality,which proves that the musical noise is suppressed and the speech signal quality is improved.(2)The combination of wavelet packet decomposition and Wiener filtering suppresses the musical noise in spectral subtraction.The wavelet packet decomposition is used to respectively process the enhanced speech signal and the clean speech signal with musical noise to obtain the wavelet packet coefficients of different frequency bands,Then the wavelet packet coefficients in each frequency band are processed by the Wiener filtering,and finally the wavelet packet coefficient replacement and the wavelet packet reconstruction are used to obtain the processed speech signal.In the simulation experiment,the enhanced speech signals in denoising and dereverberation are processed respectively.The results are compared by using the spectrogram and the perceptual evaluation of speech quality,which illustrates that the musical noise is suppressed and the speech signal quality is improved.(3)Combining beamforming to suppress musical noise in spectral subtraction.Delay compensation and spectral subtraction are performed on the speech signal of each channel in the microphone array,and the phase of the speech signal is used to suppress the musical noise in combination with beamforming.In the simulation,the denoising signal and the output result are compared by using the spectrogram and the perceptual evaluation of speech quality,which proves that the musical noise is suppressed and the speech signal quality is improved.