ctPuLSE: Close-talk, and pseudo-label based far-field, speech enhancement

Area of Science:

• Speech Processing
• Artificial Intelligence
• Machine Learning

Background:

• Current neural speech enhancement models rely on simulated data, limiting real-world performance.
• Training directly on real mixtures is challenging due to the lack of clean speech supervision.

Purpose of the Study:

• To develop a method for training speech enhancement models directly on real-recorded mixtures.
• To improve the generalizability of far-field speech enhancement models to real-world conditions.

Main Methods:

• Propose ctPuLSE: train an enhancement model on simulated data to process real close-talk mixtures.
• Utilize the enhanced close-talk speech as pseudo-labels for training far-field enhancement models on real paired mixtures.

Main Results:

• ctPuLSE effectively generates high-quality pseudo-labels from real close-talk mixtures.
• The proposed method significantly enhances the generalizability of far-field speech enhancement models on real data.

Conclusions:

• Close-talk pseudo-labeling offers a viable solution for supervised training on real-world speech data.
• ctPuLSE demonstrates a promising approach for robust neural speech enhancement in practical scenarios.