Recursively Adaptive Randomized MultiTree Coding for First Responder Voice Communications

The recent development of a non-block based speech codec that operates at 12.24 kilobits/s and lower provides the opportunity to re-examine some important speech coding applications. The new Recursively Adaptive Randomized MultiTree Coder (RAR-MTC) adapts the parameters of the code generator, that is, the parameters of the pole/zero speech model and the pitch predictor, based only upon the transmitted codec excitation and the reconstructed output speech at the Decoder. For clean, narrowband speech sampled at 8,000 samples/sec, the RAR-MTC achieves performance competitive with the well-established, block-based Adaptive Multirate speech codec for bit rates at 12.2 kbits/s and below. Since the RAR-MTC is more of a time domain waveform following codec than the popular block based parameter adaptation codecs such as AMR-NB, the RAR-MTC can reproduce the voice of emergency first responders and the background noises in their environment, such as sirens, metal saws, pneumatic chisels, and water flow, more naturally than block based codecs. This paper provides a first pass comparison of the RAR-MTC and the AMR-NB codec for emergency first responder environments. Objective performance results show that the reconstructed voice plus background noise performance is close for the two codecs, while informal listening tests indicate that the AMR-NB codec produces cleaner sounding speech but sometimes with subtle spectral distortions while the RAR-MTC reconstructed voice has a very low level hissing sound but both the voice and the background noises are very natural sounding.