Adaptive speech enhancement using frequency-specific SNR estimates

We describe an adaptive speech enhancement technique based on selecting a set of pre-computed FIR filters to process the compressed short-time power spectral trajectories of noisy speech. The responses of the pre-computed filters depend only on the signal to noise ratios (SNRs) and do not depend on the center frequency of the subbands. This allows for a compact design in which the estimate of the SNR at the particular frequency channel is used as the filter selection criterium for that sub-band.