Deep learning-based environmental source separation and sound enhancement: Advancements for cochlear implant and...

Area of Science:

• Auditory Perception
• Acoustic Signal Processing
• Hearing Impairment Rehabilitation

Background:

• Normal hearing (NH) individuals effectively associate non-linguistic sounds (NLSs) with their sources in complex environments.
• Cochlear implant (CI) users face significant difficulties in identifying and tracking NLSs, impacting autonomy and environmental awareness.
• Previous research established NLS classification methods and proposed enhancement solutions for CI and NH listeners.

Purpose of the Study:

• To introduce an experimental framework for investigating competing environmental sound perception in CI and NH listeners.
• To evaluate source separation (SS) and non-linguistic sound enhancement (SSE) techniques for improving NLS perception.
• To assess perceptual outcomes including interference, audio quality, and distortion in a two-source mixture model.

Main Methods:

• Developed a two-source mixture model with 'target' and 'interference' characteristics.
• Implemented three evaluation conditions: mixed-baseline, source separation (SS), and source separation with non-linguistic sound enhancement (SSE).
• Conducted listener evaluations using subjective ratings and forced-choice preference tests measuring interference, audio quality, and distortion.

Main Results:

• Statistically significant improvements in interference reduction were observed for CI listeners with category-matched interference and for NH listeners across all NLS categories.
• Pairwise forced-choice tests indicated a preference for SSE-processed nature and domestic noises among both CI and NH listeners.
• The study demonstrated the effectiveness of the proposed framework in evaluating SS and SSE for NLS mixtures.

Conclusions:

• The experimental framework effectively addresses challenges in competing environmental sound perception for CI and NH listeners.
• Source separation with non-linguistic sound enhancement (SSE) significantly improves perceptual outcomes, particularly for nature and domestic sounds.
• The findings provide valuable perceptual measures for characterizing the listening experience of NH and CI users in complex acoustic environments.