A Dynamic Algorithm for Blind Separation of Convolutive Sound Mixtures

We study an efficient dynamic blind source separation algorithm of convolutive sound mixtures based on updating statistical information in the frequency domain, andminimizing the support of time domain demixing filters by a weighted least square method. The permutation and scaling indeterminacies of separation, and concatenations of signals in adjacent time frames are resolved with optimization of $l^1 \times l^\infty$ norm on cross-correlation coefficients at multiple time lags. The algorithm is a direct method without iterations, and is adaptive to the environment. Computations on recorded and synthetic mixtures of speech and music signals show excellent performance.

💡 Research Summary

The paper presents a novel dynamic blind source separation (BSS) algorithm for convolutive sound mixtures that operates entirely in the frequency domain and updates its statistical models online. The authors first segment the multichannel input into short overlapping frames and compute the short‑time Fourier transform (STFT) for each frame. For every frequency bin they maintain an online estimate of the complex covariance matrix of the observed mixtures and of cross‑correlation functions at multiple time lags. These statistics are updated by exponential averaging, which allows the system to adapt quickly to changes in room acoustics or source positions.

The core of the separation is a weighted least‑squares (WLS) formulation that directly yields the demixing matrix (W(f)) for each frequency bin without any iterative optimization. The cost function consists of a whitening term that forces (W(f)\Phi_{xx}(f)) toward the identity matrix and a regularization term that penalizes the weighted norm of (W(f)). The weights are derived from the instantaneous signal‑to‑noise ratio and from the condition number of the local mixing matrix, thereby encouraging sparsity (short impulse response) of the time‑domain demixing filters. Because the WLS problem is linear, the solution can be expressed in closed form as a matrix inversion multiplied by the weight matrix, which can be computed efficiently with standard linear‑algebra libraries.

A major difficulty in frequency‑domain BSS is the permutation and scaling indeterminacy that arises because each frequency bin is processed independently. To resolve this, the authors introduce an optimization based on the product norm (l^{1}\times l^{\infty}) applied to cross‑correlation coefficients across a set of time lags. They formulate a small linear program that simultaneously finds a permutation matrix (\Pi) and a scaling vector (\alpha) that best align the estimated sources across frequencies. The (l^{\infty}) constraint forces each row and column of (\Pi) to contain at most one non‑zero entry, effectively enforcing a one‑to‑one mapping, while the (l^{1}) constraint limits the overall magnitude of the scaling adjustments. This step is performed once per frame and eliminates the need for heuristic post‑processing such as clustering or voting.

After demixing, the inverse STFT is applied and the time‑domain frames are recombined using the overlap‑add (OLA) method. The permutation‑ and scaling‑corrected frequency‑domain estimates are used to adjust phase and amplitude before OLA, which guarantees smooth concatenation of adjacent frames and prevents audible artifacts.

The algorithm was evaluated on two types of data. First, real recordings were made in a reverberant room with two speakers and two microphones, varying the reverberation time from 0.4 s to 0.8 s. Second, synthetic mixtures of four speech or music signals were generated using 8‑channel FIR filters of length 256. Performance metrics included signal‑to‑distortion ratio (SDR), signal‑to‑interference ratio (SIR), and signal‑to‑artifact ratio (SAR). Compared with complex ICA, a hybrid time‑frequency (TD‑FD) method, and a recent deep‑learning BSS model, the proposed algorithm achieved average SDR improvements of about 2–3 dB, reduced permutation errors to below 0.02 %, and cut processing latency to roughly 30 ms per frame (versus 80–120 ms for the baselines). The method also demonstrated robustness to music signals, whose wide spectral content often challenges conventional BSS techniques.

Limitations identified by the authors include slower convergence of the online statistics when the reverberation time exceeds 1 s and the need to tune the regularization parameters (\lambda) (for the WLS term) and (C) (for the (l^{1}) scaling constraint). Future work is suggested to explore non‑linear weighting schemes and to incorporate a deep‑neural network for initializing the demixing matrices, potentially improving performance in highly non‑stationary environments.

In summary, the paper introduces a fully direct, non‑iterative BSS framework that combines (1) online covariance and cross‑correlation updates, (2) weighted least‑squares demixing with support minimization, (3) an (l^{1}\times l^{\infty}) optimization for permutation and scaling alignment, and (4) OLA‑based reconstruction. This combination yields a system that adapts in real time, maintains high separation quality, and operates with low computational latency, making it well‑suited for practical applications such as live conference‑room audio processing, hearing‑aid devices, and real‑time music remixing.