The developer of the game Dungeon Cult, built on Unreal Engine 5, used local machine learning models on an RTX 5060 Ti with 16 GB of VRAM to generate sound effects and music. The project utilized Stable Audio 3 Medium for creature sounds and effects, and ACE-Step 1.5 for musical compositions.
For background music, the developer employed LoRA fine-tuning on six tracks from the game Dungeon Keeper. The training dataset consisted of 32 one-minute segments. The process took 1 hour, 10 minutes, and 6 seconds using parameters including rank 64, alpha 128, and a learning rate of 0.0001. The developer opted for a version saved after 100 epochs rather than the final one.
The background music generation involved creating a one-minute base track, then extending it by 45-second increments to achieve a 3:15 duration. This method allowed for the creation of nine total compositions. For the game's intro, a vocal track was generated using a one-minute audio reference and specific lyrics without fine-tuning, relying on the base ACE-Step 1.5 SFT model.
Creature sound effects were generated by testing three distinct character types in Stable Audio. The developer processed eight events with three variations each in approximately 94 seconds. The workflow prioritized defining the voice character first, followed by specific reactions like pain or movement, rather than generating long, complex scenes.
