Super-Gaussian non-stationary audio noises sparse representation

The super-gaussian non-stationary audio noises could not be well represented, feature extracted, and then classified in the stationary and linear transform domains as DFT (Discrete Fourier Transform) or STFT (Short Time Fourier Transform) especially in the very low SNR (Signal-to-Noise Ratio) data capturing times. DWT (Discrete Wavelet Transform) is the most commonly used and conventional transform for representing, feature extracting, and then classifying of such signals using data independent kernels. But the simulations confirm that the sparse representation transforms could well represent them than DWT because of using data dependent kernels (Atoms). In this paper it is investigated using MP-TFD (Matching-Pursuit Time-Frequency Decomposition) technique for the super-gaussian non-stationary audio noises representing, then applying NMF (non-Negative Matrix Decomposition) technique for decomposing of the TFM (Time-Frequency Matrix) into its significant components, and finally extracting MFCCs (Mel-Frequency Cepstral Coefficients) as the features in order to the sources classifying.