Introduction
Blind Source Separation (BSS), often referred to as the "cocktail party problem," involves extracting individual sound sources from a mixture without prior knowledge of the sources or the acoustic environment. Independent Vector Analysis (IVA) is a powerful technique for frequency-domain BSS that preserves the dependency between frequency bins of the same source. However, traditional IVA methods often rely on fixed source priors, such as the Laplace distribution or pre-trained Gaussian Mixture Models (GMMs). While GMMs are flexible, they typically require a training phase to capture the non-stationary nature of speech, where power levels fluctuate significantly over time.
The AV-GMM Approach
To address the limitations of static priors, this research introduces the Amplitude Variable Gaussian Mixture Model (AV-GMM). The core innovation is the introduction of a time-varying parameter, $h_t$, which scales the precision of the Gaussian components at each time frame. This allows the model to dynamically adapt to the temporal power fluctuations inherent in human speech without requiring any pre-training.
The separation process is optimized using a specialized Expectation-Maximization (EM) algorithm. This algorithm simultaneously updates the demixing matrix and the signal model parameters. To ensure stability and computational efficiency, the authors employ a unitary constraint on the demixing matrix and a pre-whitening step for the observed signals. This mathematical agility allows the model to converge significantly faster than existing flexible frameworks like NMF-IVA.
Experimental Setup
The performance of the AV-GMM-IVA was evaluated using speech segments from the TIMIT database and custom recordings in a reverberant room (8m x 6m x 3m, $T_{60} \approx 300$ms). The experiments involved 30 different mixed signals with target-to-interference ratios ranging from 6 dB to 14 dB.
As shown in the figure above, multiple source layouts were tested to ensure the robustness of the separation across different spatial configurations.
Results and Discussion
The AV-GMM-IVA was compared against three baseline methods: standard GMM-IVA (without pre-training), SSL-IVA (using a Laplace prior), and NMF-IVA.
Separation Performance
The proposed method consistently outperformed standard GMM-IVA and SSL-IVA, achieving Signal-to-Interference Ratio (SIR) improvements comparable to the high-performance NMF-IVA framework.
The results in the figure above demonstrate that AV-GMM-IVA (dark blue) maintains high separation quality across various energy ratios and spatial setups. Notably, while conventional GMM-IVA (lightest blue) fails to separate sources effectively from random initialization, the AV-GMM-IVA remains robust.
Convergence Speed
A standout feature of the AV-GMM-IVA is its convergence speed. The model reaches a steady state in fewer than 30 iterations, whereas NMF-IVA and SSL-IVA require 172 and 286 iterations, respectively. This represents a nearly 10-fold increase in speed compared to standard iterative techniques, making it highly suitable for real-time applications.
Conclusion
The AV-GMM-IVA offers a training-free, computationally efficient solution to the cocktail party problem. By incorporating a time-varying amplitude parameter, the model successfully handles the non-stationarity of speech while maintaining high separation quality and rapid convergence. Future work will focus on extending this framework to underdetermined scenarios where the number of sources exceeds the number of microphones.
Summary updated