DEEP ATTRACTOR NETWORK FOR SINGLE-MICROPHONE SPEAKER SEPARATION.

Area of Science:

• Speech processing
• Machine learning
• Signal processing

Background:

• Deep learning has advanced speech processing, but separating simultaneous speakers remains difficult.
• Key challenges include arbitrary source order and an unknown number of speakers.
• Existing methods often struggle with these complexities.

Purpose of the Study:

• To propose a novel deep learning framework for single-channel speech separation.
• To address the challenges of arbitrary source permutation and an unknown number of sources.
• To improve the accuracy and efficiency of speech separation systems.

Main Methods:

• Developed a deep learning framework utilizing attractor points in high-dimensional embedding space.
• Attractor points, derived from source centroids, group time-frequency bins of individual speakers.
• Employed end-to-end training to minimize source reconstruction error and optimize embeddings.

Main Results:

• The proposed model demonstrated effectiveness without depending on the number of sources.
• Two testing strategies, K-means and fixed attractor points, were explored.
• Achieved a 5.49% improvement over state-of-the-art methods on the Wall Street Journal dataset.

Conclusions:

• The novel deep learning framework offers a robust solution for single-channel speech separation.
• The attractor point method effectively handles arbitrary source permutation and an unknown number of speakers.
• The real-time capable fixed attractor point strategy shows significant promise for practical applications.