Comparison of a target-equalization-cancellation approach and a localization approach to source separation

Area of Science:

• Acoustics
• Signal Processing
• Auditory Perception

Background:

• Binaural cues, such as interaural differences, are crucial for auditory source localization and selective listening in complex acoustic environments.
• Existing source separation algorithms often rely on localizing time-frequency (T-F) units to enhance target sound streams.

Purpose of the Study:

• To propose and evaluate a novel binaural cue for audio source separation based on equalization-cancellation (EC).
• To compare the performance of the EC-based cue against traditional localization cues in source separation tasks.

Main Methods:

• A new source separation method using an equalization-cancellation (EC) cue is proposed.
• The effectiveness of EC is measured by the energy change in each time-frequency (T-F) unit after cancellation.
• Performance is evaluated using simulated multi-talker and diffuse-babble noise conditions.

Main Results:

• The proposed target-EC cue-based algorithm demonstrated superior source separation performance compared to localization cue-based algorithms.
• Results showed better performance in direct comparison with the ideal binary mask.
• Improved speech intelligibility was observed for target streams separated using the EC-based method.

Conclusions:

• The equalization-cancellation (EC) cue offers a more effective binaural cue for audio source separation than traditional localization cues.
• This approach shows significant potential for improving auditory scene analysis and speech intelligibility in challenging acoustic environments.