Principal component-based spectral iteration sparsity for speech enhancement
To address the inadequate performance of existing spectrum sparsity methods in speech enhance-ment under complex environments,an iterative spectrum sparsity method based on Principal Component Analysis(PCA)is proposed.First,the spectrogram of the input signal is processed by two-dimensional me-dian filtering,yielding the row component spectrum and column component spectrum.PCA is then applied to the row component spectrum sequence containing the main vocal part of the speech,to eliminate noise and preserve the main speech structure.Next,the column component spectrum sequence is combined with scaling factors for speech signal reconstruction.A dynamic scaling factor is employed to effectively control the noise present in the column component spectrum sequence.Based on this the proposed method utilizes the noise suppression effect of sparsification to perform multiple sparsifications on the spectrum to reduce noise.Experi-mental results demonstrate the improved performance,with average improvements in the signal-to-noise ratio of 13.89 dB,11.97 dB,5.65 dB,5.26 dB,and 4.73 dB,respectively,for different types of noise includ-ing White,Pink,Babble,Volvo,and Factory noise,when the input signal-to-noise ratio is set at 15 dB.Moreover,the proposed method also effectively suppresses noise while retaining important speech features across other signal-to-noise ratios,and reduces speech distortion caused by background noise.
国家科技期刊平台
NSTL
万方数据
