Discrete frame bandwidth concentration features and music recognition in complex soundscapes for cochlear implants

Abstract Current noise reduction algorithms for cochlear implants (CIs) adopt a binary speech–noise discrimination mechanism, which tends to erroneously suppress background musical signals. Meanwhile, conventional acoustic features suffer insufficient robustness in real complex soundscapes such as indoor reverberation and amplitude perturbation, severely degrading musical perception for implant recipients. To address this problem, this paper explores the inherent frame-wise distribution law of audio bandwidth per frame and proposes the acoustic feature of Discrete Frame Bandwidth Concentration (DFBC), enabling accurate identification and preservation of musical signals under complex acoustic conditions. First, spectrogram and temporal feature analyses are performed to clarify frequency-domain discrepancies among speech, music and ambient noise, verifying that musical signals possess exclusive stepwise discrete bandwidth variation characteristics. Spectral centroid trajectory typing for music is accomplished using the GTZAN dataset, which validates that the discrete concentration property of bandwidth per frame is immune to musical genres and spectral patterns, with superior stability compared with spectral centroid. A single-threshold DFBC binary classification model for music/non-music discrimination is constructed, with Mel-Frequency Cepstral Coefficients (MFCC) set as the baseline for comparative experiments under four practical CI operating conditions: amplitude fluctuation, sampling rate distortion, additive noise and indoor reverberation. The experimental results demonstrate that DFBC achieves a classification accuracy of 80.07% under clean audio conditions. This feature exhibits outstanding stability against amplitude and sampling rate variations, with further improved recognition performance in reverberant indoor environments; specifically, the accuracy reaches 84.83% at a reverberation time of 0.4 s, reflecting far stronger anti-reverberation capability than MFCC. Its performance is only limited under intense noise scenarios. Featuring clear physical interpretability and low computational complexity, the DFBC feature well satisfies the low-power operation requirements of embedded CI hardware. It can serve as a novel feature support for optimizing scene-adaptive noise reduction algorithms and facilitating high-quality auditory reconstruction in cochlear implant systems.