A Framework for Speech Activity Detection Using Adaptive Auditory Receptive Fields.

Area of Science:

• Auditory Neuroscience
• Computational Auditory Processing
• Speech Signal Processing

Background:

• The brain dynamically adapts neural processing to focus on relevant sounds amidst complex auditory environments.
• Cortical spectro-temporal receptive fields (STRFs) are known to retune adaptively, enhancing target sound features and suppressing distractors.
• Speech Activity Detection (SAD) systems struggle in noisy environments, limiting their performance in automated speech processing.

Purpose of the Study:

• To investigate task-driven adaptation in neurophysiological STRFs for speech processing.
• To demonstrate how adapted STRFs can improve speech activity detection in noisy conditions.
• To evaluate the ability of adapted STRFs to capture speech modulations for enhanced noise robustness.

Main Methods:

• Inducing task-driven adaptation in an ensemble of neurophysiological STRFs.
• Comparing the performance of adapted vs. unadapted STRF ensembles for speech activity detection in noise.
• Utilizing a stimulus reconstruction task to assess spectro-temporal modulation capture.

Main Results:

• Speech-adapted STRFs were shown to enhance speech modulations while suppressing nonspeech sounds.
• An adapted STRF ensemble significantly outperformed unadapted ensembles and a baseline in detecting speech in unseen noisy environments.
• The adapted STRF ensemble demonstrated superior capture of attended speech modulations in both clean and noisy conditions.

Conclusions:

• A biologically plausible adaptation framework can dynamically adjust feature representations in speech processing systems.
• This adaptive approach significantly improves the noise robustness of speech activity detection.
• The findings suggest a promising direction for developing more effective automated speech processing in real-world noisy scenarios.