Model based spectrum prediction

This paper presents methods for speech spectrum prediction based on Gaussian mixture models. Spectrum prediction may be useful in a packet transmission system where the sensitivity to packet losses is a major problem. Models of speech are trained by the expectation maximization algorithm using pairs, triples etc. of consecutive cepstral vectors. The models are used to design first, second etc. order predictors. The prediction schemes are evaluated using the spectral distortion criterion and compared to a simple reference method. The best prediction scheme obtains an average spectral distortion that is 0.46 dB less than for the reference method.