Yudong He

Sparse component analysis (SCA) is a popular underdetermined blind speech separation (UBSS) method. It models all sources to have an identical distribution. As speeches do not have identical distribution, SCA performs suboptimal. Some studies have improved the performance of SCA by weighting the sources through a reweighting scheme. However, they are not UBSS methods because they assume that the mixing process is known. This paper proposes a novel weighting scheme, called sparse spatial component analysis (SSCA) without the need to know the mixing process. In SSCA, weights, sources, and the parameters for modeling the mixing process are jointly optimized, making it a UBSS method. Simulation experiments show that for instantaneous mixtures, SSCA outperforms SCA and reweighted SCA, improving the source-to-distortion ratio (SDR) by 4 dB and reducing the computational time by 40%. Further, experiments using real-world recordings reveal that SSCA outperforms multichannel non-negative matrix factorization and full-rank covari ance analysis (FCA) in terms of SDR. The speed of SSCA is 200% faster than FCA.

Weighted l1 minimization schemes are common methods to achieve compressed sensing (CS). However, they fail in the presence of inaccurate prior knowledge or improper scaling of weights due to inappropriately assigned large weights causing large and destructive errors in signal recovery. This paper proposes a theory-based algorithm to identify and correct such destructive weights for each signal entry. The enhancement is achieved through a novel sparsity-inducing property (SIP) which establishes a necessary condition for successful signal recovery. SIP outperforms existing properties such as coherence, restricted isometry property, and nullspace property by indicating which signal entries fail to be recovered. This unique advantage enables us to correct destructive weights that do not satisfy the SIP condition, making signal recovery successful where it previously failed. Results from many numerical experiments demonstrate that our proposed method can improve the signal recovery capability, robustness, and stability of the weighted l1 minimization for a wide range of applications, including sparse and compressive signal recovery, noise-aware recovery, sparse error correction, fast image acquisition, and sub-Nyquist sampling.

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/demo-icassp2022/"