Gaussian mixture model based audio coding in a perceptual domain

Gaussian mixture model based vector quantization (GMM-VQ) is a powerful technique for structured vector quantization. This thesis describes the implementation and evaluation of a perceptual audio coder using GMM-VQ. It is combined with a perceptual transform using psychoacoustic preand post-filtering. The coder was tested for low rate mono audio coding (64 kbps) and the results were compared to audio encoded with MPEG-1 layer III (MP3). The evaluation was performed using a MUSHRA test, where it was demonstrated that the proposed coder could achieve results that are comparable with MP3. The coder was tested using both the psychoacoustic filter and a standard whitening LPC-filter to evaluate the difference between them. Furthermore, the benefits of frequency warping in the psychoacoustic filter was tested by comparing it to a filter using linear frequency resolution.