Research About Audio Inpainting Based On Sparse Representation

In our usual life, audio is frequently affected by the noise of environment, the congestion of network or the limited amplitude of electronic device. As a result, these boring interference will make audio subjected to impulsive noise, packet loss or clipping, which declines the audio quality in terms of audibility as well as intelligibility. Moreover, if distorted audio is applied to speech recognition or speaker identification, it will lead to recognition accuracy dropping sharply. Thus, it is necessary to restore impaired portion of audio by digital signal processing method for the purpose of improving the audio quality. According to the state-of-the-art on audio inpainting, a modified audio inpainting algorithm based on sparse representation is introduced in this thesis.The research about audio inpainting based on sparse representation refers to the restoration of impaired audio by taking advantage of reliable part’s sparse representation on over-complete dictionary under the assumption that audio is generated from sparse model. Due to the redundant information between adjacent frames in audio signal, the missing or unreliable samples can be recovered by exploiting reliable information surrounded them. In general, the audio inpainting framework based on sparse representation can be divide into two stages: the construction of over-complete dictionary and the sparse representation of signal. Considering these steps, we present two modifications in this thesis based on previous work:1. The training method of over-complete dictionaryAs for the construction of over-complete dictionary, previous articles mainly employ static dictionary. However, audio has a great variety of kinds. Different audio has different structured information such as fundamental frequency or formant. Therefore, static dictionary can not reflect structural characteristics of particular audio very well. In order to deal with this problem, we adopt training method to learn an adaptive dictionary. When choosing dictionary training algorithm with consideration about the efficiency of the algorithm and the coherence of atoms in the dictionary, we introduce the INK-SVD dictionary learning method which is a modified version of K-SVD algorithm by adding decorrelation step to it for training audio data.2. The algorithm of sparse representationPrevious researchers mostly apply OMP algorithm to get sparse representation. This method never selects overlapped atoms so that the residual error of sparse representation is very little. However, this approach does not make full use of correlative relationship between adjacent frames, in other words, there is great redundance between frames, yet we usually ignore. In this thesis, by means of experiment, we prove the existence of this correlation. In order to take advantage of this correlation information, we make an improvement of traditional OMP algorithm. When selecting atoms, we also consider whether this atom is used in the previous frame. If it is, we will increase appropriate weight of it in the next frame. The value of weight can be adjusted dynamically depend on coherence of frames. By this way, it will improve the accuracy of atoms selection and lower the computing error.In the experimental part of this thesis, we perform three experiments in accordance with three different distortion such as impulsive noise, packet loss and clipping on audio dataset provided by Audio Inpainting Toolbox. As the result shown, our algorithm improves SNR by 4 to 6 dB and PESQ by 0.2 to 0.3 compared to traditional fixed dictionary.