Linear Methods for Efficient and Fast Separation of Two Sources Recorded with a Single Microphone

Area of Science:

• Signal Processing
• Acoustics
• Machine Learning

Background:

• Monaural source separation is challenging due to overlapping acoustic signals.
• Existing nonlinear methods can be computationally intensive and complex.
• Linear methods offer a promising alternative for real-time audio processing.

Purpose of the Study:

• To evaluate three linear methods for separating two speakers from a monaural recording.
• To compare the performance of these linear methods against nonlinear approaches.
• To establish benchmark solutions for monaural source separation tasks.

Main Methods:

• Eigenmode analysis of covariance difference on sound spectrograms.
• Maximum likelihood demixing assuming Gaussian signal models.
• Suppression-regression using autoregressive models for source separation.

Main Results:

• Linear methods demonstrated competitive performance in separating male and female speakers.
• Evaluation metrics included residual error, signal-to-noise ratio, and speech quality.
• The tested algorithms showed advantages in simplicity and real-time suitability.

Conclusions:

• Linear methods provide effective and efficient solutions for monaural speaker separation.
• These approaches are simpler and more suitable for real-time applications than some nonlinear techniques.
• The study offers valuable benchmark results for monaural source separation research.