Study On Speech Enhancement Algorithms Based Real-valued Discrete Gabor Transform

In the real world, the speech signal is difficult to avoid contamination by the noise source, a major purpose of the speech enhancement from the signal with noise as much as possible the recovery of the clean speech signal. Speech Enhancement technology plays an important role in the field of speech processing speech recognition, speech coding, and human-computer voice interaction. This paper studies speech enhancement method based on Real-valued Discrete Gabor Transform (RDGT).The paper first introduces the theory of Real-valued Discrete Gabor Transform(RDGT) and speech enhancement, and derive the expression of time-frequency domain analysis and reconstruction of the speech signal through the RDGT and the corresponding inverse transform(IRDGT).Speech signal is transformed to the joint-time-frequency domain by RDGT which not the common short-term windowed Fourier transform (STFT) and also introduces some advantages of RDGT.The paper propose a novel spectral subtraction based on RDGT, the noise spectrum estimated in the joint time-frequency domain based on improved minimum statistics and the optimal smoothing algorithm of Martin. The clean speech is got by inverse transform RDGT. Experimental results show that the subjective and objective indicators are better than the traditional spectral subtraction and Martin spectral subtraction, and effective in avoiding the residual musical noise.The paper studied and proposed a novel algorithm of the optimal estimation of clean speech based on RDGT in minimum mean square error (MMSE) based log-amplitude estimator, noise spectrum estimation based on improved minimum controlled recursive averaging method (IMCRA), the experimental results showed that the effectiveness of the algorithm compared with the results of the traditional MMSE.The paper presents a novel speech enhancement algorithm. MMSE method is analyzed when the clean speech is modeled by a Laplacian distribution and the noise is modeled by a Gaussian distribution. The experimental result showed that the proposed method can achieve a more significant noise reduction and reduce the chances of speech distortion.