Gaussian mixture model based coding of speech and audio
Sam Vakil · 2004
The transmission of speech and audio over communication channels has always required speech and audio coders with reasonable search and computational complexity and good performance relative to the corresponding distortion measure. This work introduces a coding scheme which works in a perceptual auditory domain. The input high dimensional frames of audio and speech are transformed to power spectral domain, using either DFT or MDCT. The log spectral vectors are then transformed to the excitation domain. In the quantizer section the vectors are DCT transformed and decorrelated. This operation gives the possibility of using diagonal covariances in modelling the data. Finally, a GMM based VQ is performed on the vectors. In the decoder part the inverse operations are done. However, in order to prevent negative power spectrum elements due to inverse perceptual transformation in the decoder, instead of direct inversion, a Nonnegative Least Squares Algorithm has been used to switch back to frequency domain. For the sake of comparison, a reference subband based "Excitation Distortion coder" is implemented and comparing the resulting coded files showed a better performance for the proposed GMM based coder.