Generative models for sound field reconstruction.

Area of Science:

• Acoustics
• Signal Processing
• Machine Learning

Background:

• Acoustic array processing is limited by spatial sampling, restricting bandwidth and hindering high-frequency reconstruction.
• Satisfying the Nyquist criterion in the spatial domain for high frequencies is challenging in experimental acoustics.

Purpose of the Study:

• To investigate the efficacy of generative adversarial networks (GANs) for improving sound field reconstruction from experimental data.
• To determine if generative models can extend the bandwidth of reconstructed sound fields.
• To explore the application of GANs in overcoming bandwidth limitations in acoustic sensor arrays.

Main Methods:

• Three different generative adversarial models were employed to reconstruct spatial room impulse responses.
• The study focused on experimental data collected in a conventional room setting.
• The performance of GANs was evaluated based on their ability to improve spatio-temporal reconstruction.

Main Results:

• Generative adversarial networks demonstrated an improvement in sound field reconstruction accuracy.
• The models successfully recovered a portion of the sound field energy typically lost at high frequencies.
• The results indicate a potential for GANs to mitigate bandwidth limitations inherent in acoustic arrays.

Conclusions:

• Generative adversarial networks show promise for enhancing sound field reconstruction and overcoming bandwidth limitations in acoustic sensor arrays.
• The application of statistical learning models, like GANs, offers an encouraging outlook for future acoustic research.
• This approach may also benefit computational acoustics by reducing high-frequency computational demands.