Speech Denoising Using Joint Dictionary Learning And Sparse Representation

Speech is one of the most natural and convenient communication tools of mankind. On one hand it eliminates the distance barrier of communication between people; on the other hand it also improves the efficiency of human-machine interaction. However, the ubiquitous noises in real-world affect the quality of speech communication in different degree. Researching the effective speech denoising technology is particularly important and becomes a research hotspot in recent decades. Speech denoising means that attenuating the noise as much as possible without substantially degrading the speech, in order to achieve two goals that improving both the perceived quality and the intelligibility of speech. The difficulty arises from the nature of non-stationary and potentially speech-like real-world noise, thereby it is not appropriate to utilize the conventional speech denoising methods.In recent years, the deepening of the study of dictionary learning(DL) and sparse representation(SR) provides an effective solution for speech denoising. Through extracting features that can represent the clean speech and noise best and training them as dictionaries, the noisy speech then can be sparsely represented in these dictionaries and the clean speech part will be recovered. Unlike conventional denoising algorithms based on DL and SR, this thesis focus on both supervised and semi-supervised single channel speech denoising task, further digging the effectiveness of discriminative and related information between different kinds of signal for DL and SR. Regarding these information as main line, here we introduce the main work and innovation of this thesis.Firstly, we propose the Fisher criterion constraint-based discriminative joint DL algorithm for supervised speech denoising. In DL phase, the discriminative relationship between clean speech and noise is emphasized. On one hand, we add dictionary discriminative fidelity term to reduce the correlation between speech dictionary and nosie dictionary; on the other hand we add Fisher constraint term based on the different distributions of sparse coefficient between clean speech and noise, minimizing the within-class scatter as well as maximizing the between-class scatter of these distributions, thereby increasing the discrimination of SR.Secondly, we propose the supervised monaural speech denoising using complementary joint DL and SR algorithm. In DL phase, the relationships between noisy speech and clean speech, noisy speech and noise are emphasized for constraining the joint DL. Hence these relationships can be represented by the linear combination of the atomic-level mappings between dictionaries, reducing source confusion and source distortion. In denoising stage, we emphasize the complementary relationship between the effectiveness of clean speech and noise in SR. setting weights for them by considering the level of structure of signal and the input signal-to-noise ratio.Finally, the discriminative sparse non-negative matrix factorization based semi-supervised speech denoising algorithm is proposed. We emphasize the discriminative relationship between clean speech and noise while learning the noise dictionary from noisy speech in denoising stage, by adding the discriminative constrain term between noise dictionary and the known speech dictionary. Also the corresponding dictionary updating and sparse coding methods are showed in this thesis.