A two-stage deep learning algorithm for talker-independent speaker separation in reverberant conditions

Area of Science:

• Signal Processing
• Machine Learning
• Acoustics

Background:

• Speaker separation is crucial for isolating individual voices from mixed audio signals.
• Existing talker-independent methods primarily focus on anechoic conditions, limiting real-world applicability.
• Reverberation significantly degrades the performance of speaker separation systems.

Purpose of the Study:

• To investigate talker-independent speaker separation in realistic reverberant conditions.
• To propose and evaluate a novel two-stage deep computational auditory scene analysis (CASA) system for simultaneous speaker separation and dereverberation.

Main Methods:

• A two-stage deep CASA system was developed, separating reverberant utterances first, followed by dereverberation of the separated signals.
• The system was evaluated in real reverberant conditions using objective metrics for separation performance and speaker attribution accuracy.
• Comparison was made against a baseline one-stage deep CASA method and a talker-dependent system.

Main Results:

• The proposed two-stage deep CASA system significantly outperformed the baseline one-stage system in reverberant environments.
• Superior frame-level separation performance and higher accuracy in assigning separated frames to individual speakers were achieved.
• The system demonstrated successful generalization to unseen speech corpora, performing comparably to talker-dependent systems.

Conclusions:

• The two-stage deep CASA approach effectively addresses the challenges of speaker separation and dereverberation in reverberant conditions.
• This method offers a robust and generalizable solution for talker-independent speaker separation in practical acoustic environments.
• The proposed system represents a significant advancement in speech processing for complex acoustic scenarios.