Speech Enhancement Using Modulation Filtering And Distortion Control

The performance of speech communication systems is known to seriously degrade under adverse acoustic environments. The presence of noise can lead to the loss of speech intelligibility as well as to speech quality. These problems can generally put restrictions on the real applications of speech communication systems. For this reason, speech enhancement is vital for improving the performance of speech communication systems.Speech enhancement algorithms can be classified into two categories:time domain methods and transformed domain methods. In terms of time domain and transformed domain, this thesis presents a detailed review of various speech enhancement methods. Time domain methods exploit the short time stationary of speech for noise reduction. Transformed domain methods. Transformed domain methods depend on transforms, such as discrete Fourier transform (DFT), discrete cosine transform (DCT), etc. The main advantage of transformed domain methods lies in the relative ease of distinguishing and removing noise from speech. Transformed domain methods, such as spectral subtraction and minimum-mean square error short-time spectral amplitude estimator (MMSE-STSA), are carefully discussed. The strengths and weaknesses of transformed methods are evaluated.Speech signals can be faithfully represented by low frequency modulators which modulate higher frequency carriers. Low frequency modulators of speech have been shown to be the fundamental carriers of information in speech. In order to improve the speech quality by processing the modulators of speech, a novel speech enhancement algorithm using modulation filtering is proposed. The modulation filtering analysis framework consists of a filterbank, followed by subband envelop detection and frequency analysis of the subband envelopes. This thesis implements the modulation filtering analysis framework in its most straightforward form, the filterbank is implemented using the short-time Fourier transform (STFT), envelop detection is defined as the magnitude of the subband, and subband envelope frequency analysis is performed with the Fourier transform. Spectral subtraction is performed in the modulation domain. The experiment results show that the proposed method can remove the noise and suppress the musical noise effectively.Most speech enhancement methods can improve speech quality but not speech intelligibility. Speech enhancement methods introduce distortion. A positive difference between the clean and estimated spectra would signify attenuation distortion, while a negative spectral difference would signify amplification distortion. Conventional speech enhancement methods minimize the overall distortion without discriminating different speech distortions. This is one of the important reasons that why current speech enhancement methods cannot improve speech intelligibility. In order to improve the intelligibility of noisy speech, a novel speech enhancement method using distortion control is proposed. The priori SNR of noisy speech and the gain of conventional speech enhancement methods are exploited to estimate the signal distortion ration (SDR) in the frequency domain. The type of speech distortion is determined by SDR. Via tuning the gain of conventional speech enhancement methods, the speech distortion that damages the speech intelligibility can be properly controlled. The experiment results illustrate that the proposed method can improve the intelligibility of noisy speech considerably.