Noise robust speech coding at very low bit rates

We propose a new method for noise robust encoding of speech at very low bit rates. The method constitutes an extension to common speech-recognition/speech-resynthesis schemes, which have become feasible in recent years due to advances in speech recognition and artificial speech synthesis. Most such methods, however, suffer from a significant performance degradation in acoustic environments with background noise. Our proposed procedure is novel since speech enhancement capabilities are built directly into the coding paradigm. Denoising and coding are accomplished jointly by utilizing a statistical description of the parameter space of an underlying speech model (i.e. speech inventory). We conducted experiments with a dedicated speaker in acoustic environments with a signal-to-noise ratio of 10 dB. The proposed method was able to improve the perceptual quality of the encoded speech signal by 30% in PESQ measure at an average rate of just under 1.5 kbit/sec.