Abstract
We propose a new blind source separation algorithm based on mixtures of alpha-stable distributions. Complex symmetric alpha-stable distributions have been recently showed to better model audio signals in the time-frequency domain than classical Gaussian distributions thanks to their larger dynamic range. However, inference of these models is notoriously hard to perform because their probability density functions do not have a closed-form expression in general. Here, we introduce a novel method for estimating mixture of alpha-stable distributions based on characteristic function matching. We apply this to the blind estimation of binary masks in individual frequency bands from multichannel convolutive audio mixes. We show that the proposed method yields better separation performance than Gaussian-based binary-masking methods.
Introduction
Blind Source Separation (BSS), often referred to as the "cocktail party problem," involves isolating individual sound sources from a mixture recorded by multiple microphones. Traditional methods predominantly rely on Gaussian distributions to model signals. However, audio signals are inherently "heavy-tailed," characterized by high dynamic ranges and frequent peaks that Gaussian models—which assume data clusters tightly around a mean—fail to represent accurately. While complex, time-varying Gaussian models are often used as a workaround, they are sensitive to initialization and require extensive parameter estimation.
$\alpha$-stable distributions offer a theoretically superior alternative for modeling heavy-tailed signals. However, they are notoriously difficult to implement because their probability density functions (PDFs) lack closed-form expressions for most values of $\alpha$, rendering standard techniques like Maximum Likelihood Estimation (MLE) or the Expectation-Maximization (EM) algorithm inapplicable.
Methodology: Characteristic Function Matching
To overcome the lack of an analytical PDF, this paper proposes a "sketching" inspired approach to estimate the parameters of an $\alpha$-stable mixture model.
- Model Assumption: The approach assumes a mixture of $K$ sources where, at any given time-frequency point, one source is dominant (weak W-disjoint orthogonality). Each source is modeled by a symmetric $\alpha$-stable distribution.
- Characteristic Function Matching: Instead of the PDF, the authors utilize the characteristic function (the Fourier transform of the PDF), which has a simple analytical form for $\alpha$-stable distributions.
- CL-OMPR Algorithm: A greedy heuristic algorithm, Compressive Learning-Orthogonal Matching Pursuit with Replacement (CL-OMPR), is employed. It matches the empirical characteristic function of the observed data to the theoretical model by iteratively refining source components.
- Hybrid Clustering: After estimating steering vectors and $\alpha$ values, separation is performed via binary masking. Interestingly, the final clustering uses a Gaussian framework but leverages the more precise parameters derived from the $\alpha$-stable model.
Experimental Evaluation
The proposed method was evaluated on two distinct datasets to test its versatility across speech and music:
- TIMIT Database: A standard speech corpus where $K=3$ random speakers were mixed into 2 channels.
- QUASI Database: 10 musical excerpts (30s each) with stereo mixes of $K=4$ tracks (e.g., vocals, bass, drums).
Results and Discussion
The $\alpha$-stable method (CF-$\alpha$) was compared against standard EM-based Gaussian models and a purely spatial clustering method [25].
Performance on Speech (TIMIT)
For the TIMIT dataset, the CF-$\alpha$ method demonstrated a significant performance leap. It achieved a Signal-to-Distortion Ratio (SDR) of 2.70 dB, vastly outperforming the EM-Gaussian baseline of 0.54 dB. The Mixing Error Ratio (MER) also showed superior accuracy in estimating steering vectors, indicating higher robustness.
Performance on Music (QUASI)
In musical source separation, CF-$\alpha$ also led the field with an SDR of 4.11 dB, compared to 3.50 dB for the EM-Gaussian approach. The stability of the method is reflected in the competitive standard deviations across all metrics.
Key Insights
- Parsimony: Using $\alpha$-stable distributions allows for capturing wide dynamic ranges with time-invariant parameters, simplifying the model compared to complex time-varying Gaussian approaches.
- The Likelihood Paradox: CF-$\alpha$ often yielded a lower log-likelihood than the EM algorithm while producing superior separation results. This suggests that for heavy-tailed audio, maximizing Gaussian likelihood may be an inappropriate objective for source separation.
- Robustness: The method showed lower standard deviations in the Mixing Error Ratio (MER), indicating greater stability and less sensitivity to random initializations than EM-based methods.
Conclusion and Future Work
While the proposed method shows clear advantages in separation quality and stability, challenges remain. The current implementation relies on an "oracle" to solve permutation ambiguity across frequency bands. Additionally, while parameter estimation uses $\alpha$-stable math, the final clustering still employs a Gaussian approximation for efficiency. Future work will focus on autonomous permutation alignment and more computationally efficient $\alpha$-stable clustering techniques.
Summary updated