Yiming Xiao’s scientific contributions

Blind separation of multipath fading signals with impulsive interference and Gaussian noise is a very challenging issue due to multipath effects, which are often encountered in practical scenarios. Since the strong coherence among multipath signals leads to the extreme superposition in time-frequency (TF) domain, this paper proposes an iterative three-stage blind source separation (ITS-BSS) algorithm for the separation of coherent multipath signals in the presence of impulsive and Gaussian noise. Specifically, an initial estimation of mixing matrix is firstly implemented by some non-TF based algorithms. Secondly, a subspace-based TF-BSS algorithm is developed to determine the number of sources contributing at each auto-source TF point and then reconstruct corresponding sources. Thirdly, the reconstructed sources at current iteration are used to further improve the estimation accuracy of mixing matrix based on the least-squares (LS) algorithm. The last two stages are repeated by iteratively updating mixing matrix and sources until satisfied performance is achieved or a predefined number of iterations is done. Numerical results on multipath phase-shift keying (PSK) and quadrature amplitude modulation (QAM) signals plus impulsive noise under various signal-to-noise ratio (SNR) conditions are provided to demonstrate the feasibility and effectiveness of the proposed ITS-BSS algorithm.

Single-channel speech separation in time domain and frequency domain has been widely studied for voice-driven applications over the past few years. Most of previous works assume known number of speakers in advance, however, which is not easily accessible through monaural mixture in practice. In this paper, we propose a novel model of single-channel multi-speaker separation by jointly learning the time-frequency feature and the unknown number of speakers. Specifically, our model integrates the time-domain convolution encoded feature map and the frequency-domain spectrogram by attention mechanism, and the integrated features are projected into high-dimensional embedding vectors which are then clustered with deep attractor network to modify the encoded feature. Meanwhile, the number of speakers is counted by computing the Gerschgorin disks of the embedding vectors which are orthogonal for different speakers. Finally, the modified encoded feature is inverted to the sound waveform using a linear decoder. Experimental evaluation on the GRID dataset shows that the proposed method with a single model can accurately estimate the number of speakers with 96.7 % probability of success, while achieving the state-of-the-art separation results on multi-speaker mixtures in terms of scale-invariant signal-to-noise ratio improvement (SI-SNRi) and signal-to-distortion ratio improvement (SDRi).