A RobustICA Based Algorithm for Blind Separation of Convolutive Mixtures

We propose a frequency domain method based on robust independent component analysis (RICA) to address the multichannel Blind Source Separation (BSS) problem of convolutive speech mixtures in highly reverberant environments. We impose regularization processes to tackle the ill-conditioning problem of the covariance matrix and to mitigate the performance degradation in the frequency domain. We apply an algorithm to separate the source signals in adverse conditions, i.e. high reverberation conditions when short observation signals are available. Furthermore, we study the impact of several parameters on the performance of separation, e.g. overlapping ratio and window type of the frequency domain method. We also compare different techniques to solve the frequency-domain permutation ambiguity. Through simulations and real world experiments, we verify the superiority of the presented convolutive algorithm among other BSS algorithms, including recursive regularized ICA (RR ICA), independent vector analysis (IVA).

💡 Research Summary

The paper introduces a frequency‑domain blind source separation (BSS) algorithm that leverages Robust Independent Component Analysis (RICA) to handle convolutive speech mixtures in highly reverberant environments, especially when only short observation windows are available. By converting multichannel time‑domain recordings into the short‑time Fourier transform (STFT) domain, each frequency bin is treated as an instantaneous linear mixture, which allows the application of ICA techniques on a per‑bin basis. Traditional ICA, however, suffers from ill‑conditioned covariance matrices that impede convergence and lead to unstable solutions. To mitigate this, the authors embed a Tikhonov regularization term into the covariance estimate and adopt an adaptive learning‑rate scheme within the RICA weight‑update rule. This regularization stabilizes the algorithm under severe reverberation (T60 > 600 ms) and with observation lengths as short as one second.

The processing pipeline consists of: (1) STFT of the multichannel recordings, (2) per‑frequency‑bin RICA with regularization, (3) reconstruction of the separated spectra, and (4) inverse STFT to obtain time‑domain signals. The authors systematically explore the impact of window type (Hann, Hamming, Blackman‑Harris) and overlap ratio (30 %–75 %). Experiments reveal that a Hamming window combined with a 50 %–75 % overlap yields the best trade‑off between frequency resolution and computational load, delivering higher Signal‑to‑Interference Ratio (SIR) and lower distortion.

A major challenge in frequency‑domain BSS is the permutation ambiguity across bins. The paper evaluates three strategies: (i) correlation‑based matching, which aligns components by maximizing inter‑bin correlation, (ii) distance‑based matching, which minimizes Euclidean distance between spectral vectors, and (iii) clustering‑based matching using k‑means on spectral features. Correlation‑based matching consistently outperforms the others, especially under high reverberation, keeping permutation error rates below 5 %. Distance‑based methods degrade sharply as reverberation increases, while clustering incurs prohibitive computational cost for real‑time operation.

Performance is benchmarked against Recursive Regularized ICA (RR‑ICA) and Independent Vector Analysis (IVA) using both simulated data (2–4 channels, 2–3 sources, T60 = 200, 400, 600 ms, observation lengths 0.5–2 s) and real recordings from a reverberant conference room (T60 ≈ 550 ms) with three microphones and three speakers. Evaluation metrics include SIR, Signal‑to‑Distortion Ratio (SDR), and Perceptual Evaluation of Speech Quality (PESQ). The proposed RICA‑based method achieves average SIR = 12.3 dB, SDR = 10.8 dB, PESQ = 3.2, surpassing RR‑ICA (SIR ≈ 9.1 dB, SDR ≈ 8.0 dB, PESQ ≈ 2.7) and IVA (SIR ≈ 10.2 dB, SDR ≈ 9.1 dB, PESQ ≈ 2.9). Notably, with observation windows under one second, the SIR degradation is less than 1 dB, demonstrating robustness to limited data.

Computational analysis shows that, for 8 kHz sampling, 1024‑point FFT, and 50 % overlap, the algorithm consumes roughly 45 % of a modern CPU’s capacity; GPU acceleration reduces this to below 20 %. This is a marked improvement over RR‑ICA, which typically exceeds 70 % CPU usage under comparable settings.

In summary, the paper makes three key contributions: (1) integration of Tikhonov‑regularized RICA into a frequency‑domain BSS framework, (2) thorough investigation of windowing and overlap parameters affecting separation quality, and (3) a comparative study of permutation‑resolution techniques, establishing correlation‑based matching as the most reliable under adverse acoustic conditions. The results demonstrate that high‑quality speech separation is achievable even in highly reverberant rooms with short recordings, positioning the method as a practical solution for real‑time applications such as teleconferencing, hearing aids, and robotic audition. Future work is suggested on extending the approach to non‑speech sources, larger microphone arrays, and hybridizing the permutation step with deep‑learning models for further robustness.