Speech and music classification using spectrogram based statistical descriptors and extreme learning machine

This article proposes a novel feature extraction approach for speech/music classification based on generalized Gaussian distribution descriptors extracted from IIR-CQT spectrogram representation. IIR-CQT spectrogram visual representation provides superior temporal resolution at high frequencies and better spectral resolution for low frequencies compared to the conventional short-time Fourier transform analysis which provides uniform frequency resolution. Multi-level decomposition of the spectrogram image is then performed using the Nonsubsampled Contourlet Transform (NSCT) which a fully shift-invariant, multi-scale, and multi-direction expansion that can preserve the edges of the textural pattern of speech and music. The generalized Gaussian distribution (GGD) parameters are produced using maximum likelihood estimation (MLE) from the NSCT subbands to create the image feature descriptor. Chaos crow search algorithm is employed to chose the most relevant feature sub-set and to discard redundant features and finally the extreme learning machine classifier categorizes input audio segment into speech/music. The experimental results show that the proposed feature descriptor is effective and performs better compared to the existing approaches in the speech/music classification. In addition, mismatched training and testing results are also presented.