A scheme for perceptual speech and audio coding with damped sinusoids based on total least squares algorithms

The generalisation of traditional sinusoidal modelling to damped sinusoidal modelling has the advantage of a more compact and natural representation of transitional speech and audio fragments. Total Least Squares (TLS) algorithms are very efficient for the automatic calculation of the model parameters (damping, frequency, amplitude and phase). We present TLS algorithms and describe their behaviour in the time and frequency domain. By incorporating the TLS analysis in a subband scheme we are able to reduce the computational load drastically and to achieve an enhanced numerical stability. The study of the quantisation aspects of the model parameters leads us to the introduction of a scheme in which only the perceptually important components are coded.Hermus K., Verhelst W., Wambacq P., "A scheme for perceptual speech and audio coding with damped sinusoids based on total least squares algorithms", Proceedings 22nd symposium on information and communication theory in the Benelux, pp. 173-180, May 15-16, 2001, Enschede, The Netherlands.status: publishe