fluxions-ai/vui

100M parameter lightweight conversational text-to-speech model with breaths, laughter, multi-speaker dialogue, voice cloning, and streaming. Llama-based, on-device.

52
/ 100
Established

Built on a Llama-style causal transformer with ByT5 tokenization, it uses a custom audio codec (Fluac) combining Descript's DAC with Finite Scalar Quantization to reduce token rates 4x (21.5Hz vs 86Hz), enabling longer context windows and streaming synthesis via CUDA graphs. Trained on 40,000 hours of real conversational audio, it includes specialized checkpoint variants: ABRAHAM for single-speaker context-aware responses and COHOST for two-speaker dialogue synthesis. Integrates with Hugging Face (model hosting, VAD/segmentation pipelines) and torchcodec for audio encoding/decoding.

641 stars.

No Package No Dependents
Maintenance 10 / 25
Adoption 10 / 25
Maturity 15 / 25
Community 17 / 25

How are scores calculated?

Stars

641

Forks

63

Language

Python

License

MIT

Last pushed

Feb 25, 2026

Commits (30d)

0

Get this data via API

curl "https://pt-edge.onrender.com/api/v1/quality/transformers/fluxions-ai/vui"

Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.