Independent vector analysis for source separation using a mixture of gaussians prior

Area of Science:

• Signal Processing
• Machine Learning
• Acoustics

Background:

• Convolutive mixtures of acoustic signals pose challenges for source separation.
• Existing methods like Independent Vector Analysis (IVA) have limitations with diverse signal types and noise.

Purpose of the Study:

• To develop a uniform probabilistic framework for separating convolutive acoustic mixtures.
• To enhance Independent Vector Analysis (IVA) by incorporating flexible source priors and handling noise.

Main Methods:

• Developed three classes of Independent Vector Analysis (IVA) models: noiseless, online, and noisy.
• Utilized Gaussian Mixture Models (GMM) as flexible source priors.
• Employed Expectation-Maximization (EM) algorithms for parameter estimation and source separation.

Main Results:

• The proposed IVA models successfully separated mixtures of speech and music.
• The online IVA algorithm effectively handled nonstationary conditions.
• The noisy IVA model integrated denoising with source separation, achieving substantial performance gains.

Conclusions:

• The flexible source prior in IVA enables separation of diverse acoustic signals.
• The developed algorithms provide robust solutions for complex acoustic environments.
• The framework offers significant performance improvements in signal-to-interference ratio (SIR).