Signal Adaptive MBE Modeling for Low Bit Rate Speech Coding

The Multiband Excitation (MBE) model has been found to be suitable for coding of narrowband speech at rates below 4 kbps due to its ability to produce natural sounding speech with a compact set of model parameters and its noise robustness. However the fixed frame analysissynthesis framework of the MBE model is unsuitable for representing dynamic speech sounds such as stops and vocalic transitions. An alternative, the Signal Adaptive MBE model (SA-MBE), is proposed which adapts the time resolution of MBE analysis to the changing characteristics of the speech signal. Based on the transitory nature of speech either 10 or 20 ms analysis resolution is selected. The SA-MBE model is evaluated using subjective and objective methods and found to produce significantly better quality speech with improved intelligibility as compared to the fixed-frame MBE. Preliminary investigations suggest that the SA-MBE model parameters can be quantized to less than 3 kbps.