Feedback-Driven Sensory Mapping Adaptation for Robust Speech Activity Detection.

Area of Science:

• Computational Auditory Scene Analysis
• Bio-inspired Signal Processing
• Machine Learning for Audio

Background:

• Parsing complex acoustic scenes is challenging for computational systems due to data mismatch.
• The human auditory system excels at segmenting soundscapes by adapting neural representations.
• Existing data-driven audio processing systems struggle with real-world noisy conditions.

Purpose of the Study:

• To develop a robust speech activity detection system inspired by biological auditory principles.
• To mimic the brain's adaptation of neural representations for improved sound parsing.
• To address the data mismatch problem in computational audio processing.

Main Methods:

• Proposed a framework mimicking the auditory system's adaptation of neural input in high-dimensional space.
• Employed a 2-D Gabor filter bank with parameters retuned offline.
• Used feedback from statistical models to minimize misclassification risk for speech and nonspeech sounds.

Main Results:

• The adapted system demonstrated robustness to novel acoustic conditions.
• Achieved a marked reduction in equal error rates across various noisy databases.
• Showcased enhanced separability between speech and nonspeech features in a high-dimensional space.

Conclusions:

• Biological auditory system principles offer effective strategies for robust audio processing.
• Adapting neural representations is key to overcoming data mismatch in challenging acoustic environments.
• The developed framework advances the creation of intelligent audio processing systems.