An Experimental Performance Assessment of Temporal Convolutional Networks for Microphone Virtualization in a Car...

Area of Science:

• Acoustics
• Signal Processing
• Machine Learning

Background:

• Microphone virtualization aims to recreate acoustic signals at specific locations using data from other microphones.
• Realistic automotive environments present complex acoustic challenges due to engine noise, road vibrations, and cabin acoustics.

Purpose of the Study:

• To present experimental results of microphone virtualization in realistic automotive scenarios.
• To evaluate the performance of a Temporal Convolutional Network (TCN) for virtual microphone estimation.
• To investigate the trade-off between accuracy and computational complexity for the neural network.

Main Methods:

• A Temporal Convolutional Network (TCN) was designed to estimate acoustic signals at the driver's ear.
• An experimental setup was created in a B-segment car, recording acoustic data on smooth asphalt at variable speeds.
• Microphone signals were recorded in two scenarios: with and without a front passenger.

Main Results:

• The TCN demonstrated robust adaptation to different conditions when trained on both scenarios (with/without passenger).
• The system achieved good average performance in estimating virtual microphone signals.
• An investigation identified neural network parameters for accurate estimation with low computational complexity.

Conclusions:

• The TCN is a viable technique for microphone virtualization in automotive applications.
• The proposed method offers a good balance between estimation accuracy and computational efficiency.
• This technology has the potential to enhance in-car audio systems and driver communication.