Improved Convolutive and Under-Determined Blind Audio Source Separation with MRF Smoothing

Area of Science:

• Signal Processing
• Machine Learning
• Computational Acoustics

Background:

• Blind audio source separation (BASS) from noisy recordings is computationally challenging.
• Existing methods often struggle with under-determined and convolutive mixtures.
• Current algorithms frequently model spectrograms with independent Gaussian components, limiting performance.

Purpose of the Study:

• To improve the performance of expectation-maximization (EM) based blind audio source separation.
• To address limitations in modeling source spectrograms by incorporating structural priors.
• To enhance the separation of audio sources from convolutive and under-determined mixtures.

Main Methods:

• Modified the expectation-maximization (EM) algorithm for iterative estimation of mixing and source parameters.
• Introduced a Gibbs prior within the complete data likelihood function to enforce local smoothness.
• Utilized a Markov random field (MRF) to model interactions between neighboring spectrogram bins in frequency and time.
• Applied the enhanced algorithm to stereo audio source separation tasks.

Main Results:

• The proposed modifications significantly improved the efficiency of audio source separation.
• Incorporating local smoothness via Gibbs priors and MRFs enhanced the modeling of power source spectrograms.
• Simulations demonstrated the effectiveness of the enhanced EM-based algorithm on benchmark datasets.

Conclusions:

• The enhanced EM-based algorithm with local spectrogram smoothness provides a more robust approach to blind audio source separation.
• This method offers a significant improvement over traditional methods that assume independence across time and frequency bins.
• The findings are particularly relevant for complex audio separation scenarios involving noisy and convolutive mixtures.