Challenge #2

Generative Sound Production for a Submarine Serious Game

Context

Simulation environments play an important role in underwater acoustics research by enabling controlled experimentation and algorithm development without the logistical and financial burdens of field trials. While modern simulation frameworks incorporate increasingly sophisticated models of acoustic propagation, the source signals driving these often lack realism and are typically limited to overly simplistic mathematical models or looped recordings. This contrasts with real underwater acoustic sources, both biological (e.g., marine mammals) and anthropogenic (e.g., ships and machinery), which produce rich, non‑stationary, and non‑repeating temporal structures that vary with behaviour and environmental and operating conditions.

This problem is particularly acute in the context of a sonar serious game designed to investigate human-AI teaming (see photo below), where human operators and AI agents interact with simulated acoustic scenes. In this setting, perceptual realism is essential, because unrealistic sounds can affect operator perception and behaviour and ultimately undermine the ecological validity of experimental findings.

Objectives

Participants are invited to develop methods for synthesising realistic underwater acoustic signals from marine sources such as ships and marine mammals. The resulting audio should be suitable for integration into acoustic propagation models and broader simulation environments. The aim is not simply to reproduce or replay recorded sounds, but to generate source signals whose spectral, temporal, and modulation characteristics vary and evolve in plausible ways in response to underlying behavioural, operational, and contextual factors, for example:

• Ships
  • Variability across different classes (fishing vessel, cargo vessel, etc.)
  • Evolution with speed and manoeuvring
• Marine mammals
  • Variability across different species
  • Evolution with behaviour (socialising, hunting, etc.)

Suitable approaches may include deep generative machine‑learning, physics‑inspired or parametric models, or hybrid methods that combine both data‑driven learning with domain knowledge of sound generation mechanisms.

Resources

Participants are encouraged to make use of open access marine acoustic data for training, validation, and benchmarking.

https://acoustics.ac.uk/open-access-underwater-acoustics-data/

A propagation model will be provided to assess how generated source signals behave once placed into simulated acoustic environments. There will also be an opportunity to incorporate the synthesised sounds into the serious sonar game.