LuxTTS

Apache-2.0 voice-cloning TTS model from ysharma3501 (Yatharth Sharma), 4,867★ / 634 forks, created 23 January 2026, last pushed 5 June. Weights and a demo Space on Hugging Face (YatharthS/LuxTTS), plus a Colab. T1 — the repository and model card read directly. Its headline claims carry no benchmark, which is the main thing this page has to record.

What it actually is

A distillation of ZipVoice (k2-fsa), not a new architecture. The author states the delta plainly: “LuxTTS uses the same architecture but distilled to 4 steps with an improved sampling technique. It also uses a custom 48khz vocoder instead of the default 24khz version.” The vocoder is Vocos.

That makes this the spoke’s first flow-matching TTS source — the branch of the architecture axis (text-to-speech) that the corpus names and had no instance of, alongside the autoregressive (orpheus), neural-codec (misotts) and lightweight (kokoro) families it already holds. The distillation-to-few-steps move is what makes flow matching fast: num_steps=4 is the shipped default, and the README’s own guidance is that 3–4 is the efficiency sweet spot.

The efficiency claims

• 150× realtime on a single GPU, and faster than realtime on CPU.
• Fits in 1 GB VRAM — “it can fit in any local gpu.”
• 48 kHz output, against the 24 kHz most of the field ships.
• Currently float32; the author expects fp16 to be “almost 2x” faster and has not shipped it.

Sampling is exposed rather than hidden: t_shift (higher sounds better, worse WER), num_steps, speed, rms, return_smooth (fixes metallic artifacts), and ref_duration (lower speeds up inference; raise to 1000 if artifacts appear). A model that hands you the quality/speed knobs and documents which direction each one hurts is more useful than one that picks for you.

Cloning needs a 3-second reference clip minimum.

The claim with nothing behind it

“SOTA voice cloning on par with models 10x larger.” No MOS, no WER, no Elo, no comparison table, no eval set, no named model it was measured against. This spoke keeps tts-benchmarks, mean-opinion-score and tts-arena-leaderboard precisely for claims of this shape, and none of them has an entry for LuxTTS.

What makes it notable rather than routine: the README uses WER as a tuning axis — t_shift trades quality against WER — so the author is working with the metric and publishing no value of it. The 48 kHz output is the one differentiator that needs no benchmark, since it is a property of the vocoder rather than a quality judgement.

Set against kokoro, the efficiency leader this corpus already holds at 82M parameters: LuxTTS competes on a different quantity (throughput and VRAM ceiling rather than parameter count) and does the thing Kokoro doesn’t — clone a voice from seconds of audio.

Adoption, which is the evidence that does exist

Five independent community projects are listed by the author: a Gradio app, a ComfyUI node pack, an ONNX port, a clean-UI wrapper, and hosting on fal.ai. Third-party ports and a commercial host are weak evidence of quality and real evidence of use — the same shape of signal the spoke accepted for deployment claims elsewhere, and worth more here than the star count.

No consent language anywhere

The README carries no ethics note, no consent requirement, no watermarking, and no acceptable-use statement. It is an Apache-2.0 model that clones a voice from three seconds of audio at 150× realtime in 1 GB of VRAM, and the only guidance on the reference clip is that shorter than three seconds works badly.

Recorded because this spoke made it a standing axis: audio-deepfake documents CEO-voice fraud, robocall impersonation and non-consensual cloning, alongside ASVspoof detection and SynthID watermarking. Note the asymmetry — watermarking is a thing large labs do to their own outputs, and a distilled community model with a swapped vocoder inherits none of it. The permissive-licence trend this spoke has been tracking as good news for openness has this as its other face: the cheaper and freer cloning gets, the less any of the detection infrastructure is attached to it.

Freshness

Last pushed 5 June 2026, read 29 July. Two roadmap items remain open — LuxTTS v1.5 and the fp16 inference code — and the star count is moving faster than the repository. Treat the performance numbers as float32-era.

Related

text-to-speech · voice-cloning · open-weight-tts · kokoro · audio-deepfake · tts-benchmarks · neural-audio-codec · speech-audio-ai · synthesis