Can all variations within the unified mask-based beamformer framework achieve identical peak extraction performance?

2024-07-22

EURASIP Journal on Audio, Speech, and Music Processing

AI Summary

This work investigates whether variants of mask-based beamforming (BF) exhibit peak extraction performance equivalence under a unified masked BF framework. Addressing limitations in prior studies—namely, incomplete BF coverage (e.g., exclusion of MVDR) and reliance on ideal scaling (IS)—we propose the first comprehensive, differentiable, and realistic unified framework encompassing all major BF types, including MVDR, with mask-driven differentiable scaling. For the first time, we theoretically and empirically demonstrate that mask design (e.g., IRM, IBM, PSF), time-frequency modeling, and the coupling mechanism between masking and BF critically affect extraction consistency, thereby challenging the common assumption of performance equivalence. Evaluations on WSJ0-2mix and LibriMix reveal up to 12.7 dB disparity in peak extraction error across mask–BF combinations, confirming non-equivalence. Our core contribution is a differentiable, general, and practically viable paradigm for joint mask–BF optimization.

Technology Category

Search and Optimization: Mixed Discrete/Continuous SearchMachine Learning: Hardware-aware MLConstraint Satisfaction and Optimization: Mixed Discrete/Continuous Optimization

Application Category

User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendationSearch and Retrieval-Augmented AI: Efficiency and scalability of Web search enginesSecurity and Privacy: Large-scale security measurements

Problem

• Unified mask-based beamformer framework performance
• Optimal mask variations for target extraction
• Realistic scaling in sound extraction systems

Innovation

• Unified mask-based beamformer framework
• Filter estimation for all BFs
• Realistic scenario scaling with masks