A Theoretical Framework for Envelope-Based Time-Frequency Masks in Cochlear Implant Noise Reduction Applications

This work presents a theoretical framework for designing envelope-based time-frequency masks (TFM) in cochlear implant (CI) noise-reduction applications. A cost function is proposed with the aim of achieving the ideal envelope of the desired speech from an estimate of the noise-corrupted speech envelope. Real-world envelope estimation inaccuracies arising from the truncation of the discrete Hilbert filter are explicitly considered. Two new optimal TFMs were derived. They reduce to the square root Wiener (SRW) and Chiea and Costa (C2F) masks under ideal conditions, providing a unifying theoretical framework. Objective metrics and psychoacoustic experiments with ten normal-hearing (NH) subjects and three CI users were employed to evaluate the performance of several TFMs under near real-world conditions, considering limitations in envelope estimation accuracy and processing latency. The proposed TFMs achieve lower mean envelope errors and improve objective intelligibility scores by up to 3.4% compared to SRW and C2F, where the maximum improvement occurred at a signal-to-noise ratio (SNR) of −5 dB. Psychoacoustic experiments reveal intelligibility gains of up to 6.7% for CI users, where the utmost betterment occurred at an SNR = 0 dB. The proposed theoretical framework led to optimal envelope-based TFMs that converge to the classic SRW and C2F methods under ideal signal envelope estimation. This work provides a compelling method for designers of high-performance noise reduction systems in CIs, potentially improving real-world intelligibility.