Enhancing music recognition using deep learning-powered source separation technology for cochlear implant users.

This article focuses on enhancing music listening experiences for cochlear implant (CI) users through a deep learning-based source separation system with a self-fitting function. The study employed the Demucs_augment model, which integrates noise augmentation during training, to separate music into components (vocals, rhythm, harmony) and allows users to customize the mixing proportions according to personal preference. Objective evaluations using source-to-distortion ratio (SDR), source-to-interference ratio (SIR), and source-to-artifact ratio (SAR) demonstrated that the proposed system outperformed the baseline Demucs model, particularly in noisy environments. Subjective listening tests with vocoder simulations further confirmed that personalized self-fitting significantly improved music appreciation compared to fixed vocal-to-instrument ratios. While computational complexity currently limits real-time implementation on CI devices, the study suggests that the proposed approach holds promise for future clinical application to improve music perception among CI users.