Bachelor thesis / UE5 + Python

AI Audio Research

I tested whether AI sound effects can react fast enough inside a playable game.

Short answer: raw generation was too late. The useful version prepares sound before the player needs it.

Unreal Engine AI audio prototype with Python audio server running
Prototype Unreal sends gameplay requests to a Python audio server.

A playable UE5 prototype with a Python audio backend.

I built a UE5 client that asks a Python backend for sound during gameplay. The backend loads models, checks cache, returns PCM, and Unreal plays the result.

Python AI audio prototype terminal
First prototype: type a request, generate sound, log the timing.

The sound could be good and still be wrong.

For gameplay feedback, timing matters more than polish. If the sound arrives after the input, the player hears the delay first.

Playable window 100-150ms
Raw AI generation 1.4-3.5s

Prepare the sound before gameplay needs it.

The practical system is simple: predict likely sounds, generate them off the critical path, cache the useful results, and keep procedural DSP ready as fallback.

predict generate early cache play from cache

Same prompt, different method.

These samples come from the presentation deck. Use them as listening evidence, not as another data table.

Procedural DSP presentation visual
Procedural DSP: fast, synthetic, useful as fallback.
Procedural DSP Heavy footstep on wet stone

Low latency, but clearly synthetic.

Latency: 15-25ms Use: Fallback

Two recordings for context.

Unreal client: gameplay requests and playback.

Python backend: model load, cache hits, retries.

Bachelor Thesis · 68 Pages

Read the Full Paper

Benchmarks, cache design, Unreal implementation notes, and failure handling.

Read AI Audio Research Paper