Deflationary Extraction Transformer for Speech Separation with Unknown Number of Talkers

Area of Science:

• Speech processing
• Machine learning
• Signal processing

Background:

• Traditional speech separation often requires prior knowledge of the number of speakers.
• This limitation hinders real-world applications where speaker count is uncertain.

Purpose of the Study:

• To develop an automated speech separation method that does not require knowing the number of talkers.
• To improve the performance of speech separation in unconstrained acoustic environments.

Main Methods:

• A novel speech separation technique employing a deflationary extraction of individual talker voices.
• Utilizes a transformer-based backbone with permutation-invariant training for robust speaker identification.
• Incorporates a predefined termination criterion to automatically determine the number of speakers.

Main Results:

• The proposed method achieves superior performance compared to state-of-the-art models on Libri5Mix and Libri10Mix datasets.
• Demonstrates significant improvements even when the number of talkers is not provided as input.
• Effectively separates speech without prior knowledge of the talker count.

Conclusions:

• The developed method offers a practical solution for speech separation in scenarios with unknown speaker numbers.
• Advances the field of audio signal processing by enabling robust separation without explicit speaker count information.
• Highlights the potential of transformer architectures and permutation-invariant training for complex audio tasks.