Causal Deep CASA for Monaural Talker-Independent Speaker Separation

Area of Science:

• Speech Processing
• Computational Auditory Scene Analysis (CASA)
• Deep Learning

Background:

• Talker-independent monaural speaker separation addresses separating concurrent speakers from single-microphone recordings.
• Deep CASA, inspired by human auditory scene analysis, achieves state-of-the-art results for 2-3 speaker mixtures but is non-causal.
• Non-causal systems are unsuitable for real-time applications like telecommunication and hearing prostheses, which require causal processing.

Purpose of the Study:

• To develop a causal version of deep CASA for real-time speaker separation.
• To create a speaker-number-independent system generalizable to various numbers of concurrent speakers (C >= 2).
• To enable speaker separation in applications demanding causal processing.

Main Methods:

• Modified temporal connections, normalization, and clustering algorithms within the deep CASA framework to eliminate future information usage.
• Developed a C-speaker deep CASA system trained in a speaker-number-independent manner.
• Ensured no future information is used throughout the deep network for causal processing.

Main Results:

• The proposed causal deep CASA system effectively separates concurrent speakers from single-channel audio.
• The system demonstrates excellent performance regardless of whether the number of speakers is known or unknown beforehand.
• Achieved high-quality speaker separation performance in experiments.

Conclusions:

• The causal deep CASA approach successfully addresses the non-causal limitation of previous deep CASA systems.
• This causal system is suitable for real-time speech processing applications requiring speaker separation.
• The speaker-number-independent design enhances the system's generalizability and practical utility.