Variable bit-rate sinusoidal transform coding using variable order spectral estimation

Sinusoidal transform coding (STC) is known to be capable of producing good communication quality speech coded at bitrates below 4kb/s. Discrete all-pole modelling (DAP) is an alternative spectral estimation method which can be more accurate than the conventional linear prediction (LP) analysis normally used by STC. In the quest to achieve the highest possible speech quality at lower and lower average bit-rates in variable bit-rate coding schemes, more and more effort must be made to investigate ways of varying the number of parameters according to the characteristics of each speech frame. This paper considers the advantage to be gained by varying the all-pole model order according to the discrete Itakura-Saito (IS) distance measure used in DAP. A significant reduction is achieved in the average number of parameters to be quantised compared to the fixed order model while the speech quality remains the same.