Harvesting Partially-Disjoint Time-Frequency Information for Improving Degenerate Unmixing Estimation Technique

Yudong He, He Wang, Qifeng Chen, Richard H. Y. SoICASSP 2022 · OpenReview

Abstract

The degenerate unmixing estimation technique (DUET) is one of the most efficient blind source separation algorithms tackling the challenging situation when the number of sources exceeds the number of microphones. However, as a time-frequency mask-based method, DUET erroneously results in interference components retention when source signals overlap each other in both frequency and time domains. In this paper, to avoid the erroneous retention, instead of masking, we propose to use multiple linear spatial filters (e.g., the minimum variance distortionless response filter) to extract the desired signals. These filters are constructed based on the information embedded in the detected single-source-points, that is, time-frequency points contributed by a single source. In comparison with the conventional DUET, our method achieved an impressive improvement greater than 5 dB in the source-to-interference ratio and 2 to 5 dB improvement in the source-to-distortion ratio, respectively. Findings are substantiated by unmixing simulation using live-recorded mixture signals from up to four sources. Audio examples can be found on the web page: "https://ydcnanhe.github.io/demoicassp2022/"

Introduction

Blind Source Separation (BSS), commonly known as the "cocktail party problem," involves isolating individual sound sources from a mixture. While many algorithms require at least as many microphones as sources, real-world scenarios are frequently "underdetermined," featuring more speakers than sensors. The Degenerate Unmixing Estimation Technique (DUET) is a classic solution that relies on the assumption of "W-disjoint orthogonality," which posits that only one source is active at any specific time-frequency point. However, in practice, sources often overlap, causing DUET’s binary masks to erroneously retain interference, leading to "musical noise" and poor separation quality.

Methodology

To overcome the limitations of binary masking, we propose a three-step framework that utilizes linear spatial filtering (LSF) instead of simple 1/0 masking:

  1. Single-Source-Point (SSP) Detection: The algorithm identifies specific time-frequency points where only one source is dominant, rather than assuming disjointness everywhere.
  2. Parameter Estimation: These "clean" SSPs are used to estimate the spatial characteristics (Acoustic Transfer Function or ATF) and interference patterns for each source.
  3. Linear Spatial Filtering: We apply adaptive filters to extract target signals. We specifically highlight the Interference Suppression Response (ISR) filter. Unlike the standard MVDR filter, the ISR filter is "ATF-independent," focusing on minimizing interference without needing a perfect model of the room's acoustics, which makes it highly robust in reverberant environments.

Experimental Results

The proposed methods were evaluated using live-recorded mixtures from the SiSEC2011 dataset in environments with reverberation times of 130ms and 250ms. We compared DUET-MVDR and DUET-ISR against standard DUET and state-of-the-art methods like ILRMA and MULTINMF.

Performance in Determined and Underdetermined Scenarios

In the "determined" case (2 sources, 2 microphones), the DUET-ISR method (represented by orange bars in the figures) significantly outperformed existing techniques across all metrics, including Source-to-Distortion Ratio (SDR), Source-to-Interference Ratio (SIR), and Source-to-Artifacts Ratio (SAR).

Performance comparison for 2 sources. DUET-ISR (orange) shows superior performance across SDR, SIR, and SAR metrics.

In "underdetermined" scenarios involving 3 or 4 sources, the advantage of the ISR filter remained clear. Even as the complexity of the mixture increased, the linear filtering approach suppressed overlapping interference more effectively than binary masking.

Performance comparison for 3 sources. In the underdetermined case, DUET-ISR maintains a significant lead in SDR.

Key findings include an improvement of more than 5 dB in SIR and a 2 to 5 dB improvement in SDR compared to conventional DUET. The ISR filter's ATF-independence proved crucial in maintaining performance in reverberant conditions where precise path estimation is difficult.

Conclusion

By shifting from binary masking to adaptive linear spatial filtering based on single-source-point detection, we significantly improved the performance of the DUET algorithm. While the ISR filter introduces a mild high-pass effect, the trade-off for substantially cleaner source separation is highly beneficial for practical applications. Future work will explore minimizing this distortion and extending the method to larger microphone arrays.

Summary updated

Abstract

The degenerate unmixing estimation technique (DUET) is one of the most efficient blind source separation algorithms tackling the challenging situation when the number of sources exceeds the number of microphones. However, as a time-frequency mask-based method, DUET erroneously results in interference components retention when source signals overlap each other in both frequency and time domains. In this paper, to avoid the erroneous retention, instead of masking, we propose to use multiple linear spatial filters (e.g., the minimum variance distortionless response filter) to extract the desired signals. These filters are constructed based on the information embedded in the detected single-source-points, that is, time-frequency points contributed by a single source. In comparison with the conventional DUET, our method achieved an impressive improvement greater than 5 dB in the source-to-interference ratio and 2 to 5 dB improvement in the source-to-distortion ratio, respectively. Findings are substantiated by unmixing simulation using live-recorded mixture signals from up to four sources. Audio examples can be found on the web page: "https://ydcnanhe.github.io/demoicassp2022/"
BibTeX
@inproceedings{he2022harvesting,
  author = {Yudong He and He Wang and Qifeng Chen and Richard H. Y. So},
  title = {{Harvesting Partially-Disjoint Time-Frequency Information for Improving Degenerate Unmixing Estimation Technique}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2022},
  url = {https://openreview.net/forum?id=tGC18JMXG8J},
}
Figure