Problem: Automatic classification of the acoustical situation, and fast prediction of the local signal-to-noise ratio (SNR)

Application: VAD for mobile communication, noise suppression for e.g. hearing instruments

Motivation: Humans can easily detect and classify different sound sources, e.g., distinguish between speech and noise. Which features in the acoustic waveform allow for such impressive skills?

Approach: Modeling neurophysiological findings on amplitude modulation processing in the auditory system of mammals yielding spectro-temporal feature patterns. Classification and SNR prediction with artificial neural networks.

Implementation: Narrow-band SNR estimation fed into Wiener Filter like noise reduction algorithm. No assumption about stationarity of the noise and no speech pause detection necessary.

Paper free to download:
Tchorz, J., Kleinschmidt, M., and Kollmeier, B.: 'Noise suppression based on neurophysiologically motivated SNR estimation for robust speech recognition', Proceedings of NIPS 2000, in press. Download (zipped ps, 110 kbyte)

Audio Demo:
 

speech & drilling machine	Original	Processed by AMS noise reduction
speech & printing machine	Original	Processed by AMS noise reduction
speech & icra7 noise (speech like modulated noise)	Original	Processed by AMS noise reduction
speech & pulsed white noise	Original	Processed by AMS noise reduction