The results indicated that a single-microphone NR system based on time-frequency gain manipulation improved the HINT thresholds of listeners. However, to obtain benefit in speech intelligibility, the detectors used in such a strategy were required to detect an unrealistically high percentage of the speech energy and to perform the gain manipulations on a fast temporal basis.

A noise-reduction algorithm was developed based on a neural model for detection that is robust in fluctuating noise. The phase-opponency (PO) neural model correlates the outputs of two different auditory filters that differ in phase by 180 degrees at the target frequency. The PO detector used here consists of a pair of overlapping bandpass filters with phase responses that differ by 180 degrees near the center frequency (CF). The correlation between the filter responses is reduced when a narrowband signal near CF is present in a noisy background. A bank of PO detectors was used to process speech corrupted by additive Gaussian noise. The time-varying outputs of the detectors can be post-processed to retain information-rich regions, such as formants and frication onsets, while greatly reducing noise between formants. The final detector output controlled the gains in a separate analysis/synthesis filterbank. Spectrograms of speech sounds before and after noise reduction illustrate the ability of the system to detect major features in speech down to low signal-to-noise ratios. The quality of the processed signal will be demonstrated. This system is intended as a front-end for speech recognition systems or in hearing-aid noise-reduction algorithms. [Work supported by NSF BCS0236707 (OD,CYE) and NIH DC-001641 (MCA,LHC).]