Evaluation of a deep-learning method for implementing wide dynamic range compression and noise reduction simultaneously

Most hearing aids incorporate multi-channel wide dynamic range amplitude compression (WDRC) to compensate for the reduced dynamic range associated with sensorineural hearing loss. In theory, fast-acting WDRC should be used, because the reduced dynamic range of hearing is thought to be largely caused by reduced fast-acting compression in the cochlea. However, fast-acting WDRC has undesirable side effects, including “cross-modulation”: two sound sources (e.g., speech and noise) that are independently amplitude modulated have envelopes that are partially correlated after WDRC is applied. This hinders the perceptual separation of the sources. A deep-learning method, NN-WDRC, is described in which the speech and noise are estimated separately, fast compression is applied to the speech, and slow compression to the noise. The compressed signals are combined with a controllable amount of noise reduction. The whole system was implemented in a low-complexity network, which was trained using many talkers, audiograms, and types of background noise; the prescribed gains and compression ratios for each audiogram were based on the CAM2 fitting method. Technical measures and evaluations using human listeners indicated that NN-WDRC performed better than conventional fast-acting and slow-acting WDRC, and than “signal-to-noise ratio aware” WDRC, especially for non-stationary noises like clicking sounds and a siren.