WaveNet-based approximation of a cochlear filtering and hair cell transduction model.

Area of Science:

• Auditory Neuroscience
• Signal Processing
• Machine Learning

Background:

• Computational auditory models offer insights into hearing but are computationally intensive.
• Accurate auditory models are crucial for developing bio-inspired speech and audio processing algorithms.
• Efficient execution is often required, posing a challenge for complex models.

Purpose of the Study:

• To develop a computationally efficient approximation of the cochlear filtering and inner hair cell (IHC) transduction stages of a widely used auditory model.
• To leverage WaveNet for approximating key auditory functions.
• To enable faster and more accessible auditory modeling for research and applications.

Main Methods:

• A WaveNet model was trained to approximate the normal-hearing cochlear filtering and IHC transduction stages.
• Training utilized a large dataset of diverse audio signals (speech, music) across various sound pressure levels (SPLs) and frequencies.
• The model was evaluated on unseen noisy speech, music, sine tones, and click signals.

Main Results:

• The WaveNet model accurately predicts IHC receptor potentials for various stimuli.
• Achieved processing times up to 250 times faster than an optimized reference implementation.
• Demonstrated high accuracy across a range of SPLs (30-100 dB) and characteristic frequencies (125 Hz - 8 kHz).

Conclusions:

• WaveNet provides an efficient and accurate approximation of critical auditory model components.
• The model's speed and accuracy facilitate applications in real-time bio-inspired audio processing.
• Its differentiability makes it suitable for deep learning-based speech and audio enhancement algorithms.