Based on the fact that in every day classification tasks normal- hearing human listeners still outperform any machine-based approach, it is worthwhile to have a closer look at recent findings of other research groups in psychoacoustics and neurobiology. In this talk I will give two examples of how this knowledge can be successfully utilized to build better algorithms for automatic signal classification and automatic speech recognition.

a) Automatic short-term signal-to-noise-ratio (SNR) estimation for noise suppression This approach is motivated by neurobiological evidence of "periodotopic" gradients in the higher stages of the auditory system with respect to different modulation frequencies and the success of psychoacoustic models using a modulation-filter bank. Based on the so-called amplitude modulation spectrogram of the input signal (a spectro-temporal representation) the short-term SNR between speech and noise can be estimated by an articifical neural network. A frequency-channel-specific estimate allows for a noise reduction scheme which is especially suited for stationary types of noise.

b) Robust front end for automatic speech recognition The receptive fields of cortical neurons, on the one hand, and psychoacoustic "early auditory features" on the other hand, consist of similar patterns in the spectro-temporal representation. These are,in turn,astonishingly similar to the "secondary features" introduced by Tino Gramss a few years ago in his Feature Finding Neural Network for speech recognition. A new robust front end can be designed by extracting certain types of secondary features from the output of an auditory perception model.