TinyAudio

Checkpoints for TinyAudio: Compact and Efficient Text-to-Audio Generation for Low-Resource Deployment.

Paper · Code · Demo

TinyAudio is a complete text-to-audio system with only 87M parameters, covering the text encoder, the generator and the decoder. It generates 44.1 kHz audio with 0.48 GB of peak GPU memory. Its one-step variant, TinyAudio-MF, generates audio in real time on a four-core CPU quota.

Files

Path	Description
weights/tinyaudio.pth	TinyAudio generator, flow matching, 25 steps
weights/tinyaudio_mf.pth	TinyAudio-MF generator, MeanFlow, 1 step
weights/ta_clap_ettin32m.pt	TA-CLAP text encoder (32M)
weights/ta_vae/	TA-VAE autoencoder
cpu/standard/	ONNX export of TinyAudio-MF for CPU inference with ONNX Runtime

Usage

Checkpoints are downloaded automatically by the code repository.

git clone https://github.com/junxi25liu/TinyAudio.git
cd TinyAudio
pip install -e .
python infer.py --prompt "A dog is barking while a car passes by"

from tinyaudio import TinyAudioPipeline

pipe = TinyAudioPipeline.from_pretrained("tinyaudio_mf")
audio = pipe("A dog is barking while a car passes by")
pipe.save(audio[0], "dog.wav")

For CPU inference without PyTorch, run python cpu/tinyaudio_onnx.py "your prompt". See the code repository for training and evaluation.

Citation

@article{liu2026tinyaudio,
  title   = {TinyAudio: Compact and Efficient Text-to-Audio Generation for Low-Resource Deployment},
  author  = {Liu, Junxi and Li, Xiquan and Guan, Wenhao and Duan, Yifan and Niu, Zhikang and Huo, Yanru and Ma, Ziyang and Chen, Xie},
  journal = {arXiv preprint arXiv:2609.31525},
  year    = {2026}
}