ZappY-AI/audiogen-toycar-lora

AudioGen-medium LoRA for DCASE ToyCar normal sounds

LoRA adapter (rank 16) and a learned <toycar> text token for facebook/audiogen-medium, trained on the 1000 normal training clips of the DCASE 2026 Challenge Task 2 ToyCar machine (near-minus-far residual audio, 16 kHz, full 10 s clips). Used to generate synthetic ToyCar normal sounds for domain-adversarial training experiments.

Files:

• lora_r16_full10.pt: LoRA weights (peft inject_adapter_in_model, r=16, alpha=16) on the language model's out_proj, linear1 and linear2 layers
• toycar_token_lora_r16_full10.pt: input embedding of the added T5 token <toycar>

Generation (prompt the sound of <toycar>, cfg 3, top-k 250, 10 s): see tools/colab_gen_synth.py in https://github.com/zappy586/dcase-experiments

License follows the base model's weights (CC-BY-NC 4.0).