How to Run SupraTTS-0.1-Beta Locally: A Tiny Glow-TTS Voice Model

SupraTTS-0.1-Beta is a tiny open-weight English text-to-speech model from SupraLabs, released September 27, 2026 under the MIT license. It pairs a 29.6 million parameter Glow-TTS acoustic model with a 14 million parameter HiFi-GAN vocoder, trained from scratch on the single-speaker LJSpeech dataset. It is the successor to LH-Tech-AI's Flare-TTS-v1.5, built mainly to fix a training bug and sound more natural, and it ships with a working local inference script.

Written by Priya Raghunathan, local-AI hardware reviewer. I check parameter counts and licensing against the primary source before recommending an install path, and I flag it plainly when a linked repo is not the one the install commands actually use.

A top-20 entrant on Hugging Face's trending list

SupraTTS-0.1-Beta entered the top 20 of Hugging Face's text-to-speech trending list. At time of writing it has 8 likes and 0 recorded downloads on the SupraLabs/SupraTTS-0.1-Beta model page, which fits a model that was only just published. There is a demo Space, SupraLabs/supratts-0-1-beta (3 likes), if you want to hear a sample before downloading anything.

Specs at a glance

What changed from Flare-TTS-v1.5

SupraLabs describes this as a from-scratch retrain of the same Glow-TTS encoder and decoder Flare-TTS-v1.5 used, on the same LJSpeech dataset, with one goal: make it sound better. The model card lists the fixes plainly, including a basic one, the previous model's learning-rate scheduler was stepping once per epoch instead of once per step, so v1.5 never actually reached its target learning rate during training.

• Fixed the per-epoch learning-rate scheduler bug that kept v1.5 under-trained
• Added blank tokens between input tokens
• Swapped the deterministic duration predictor for a stochastic one borrowed from VITS, so timing varies slightly between runs instead of repeating identically
• Trained in bf16, with a Triton kernel (Super-MAS) for monotonic alignment search
• Trained the HiFi-GAN vocoder longer and from scratch on ground-truth mel spectrograms

Benchmark results

SupraLabs evaluated both models on 10 out-of-domain sentences plus 10 held-out LJSpeech sentences, across 5 random seeds. UTMOS is an automatic mean-opinion-score estimate from 1 to 5; F0-Std measures how much pitch varies within a sentence, in semitones; WER is Whisper's word error rate on the generated audio, used as a mispronunciation check.

UTMOS climbs 0.53 points over v1.5, and the pacing problem is largely gone: v1.5 ran noticeably fast at 1.087x the original speech rate, where SupraTTS-0.1-Beta lands at 1.004x, close to natural pace. Word error rate is roughly flat between the two. Pitch variation, F0-Std, is still well below real recordings on both models; SupraLabs's own notes call the new model's pitch stable but flatter than a real speaker, and name that the next thing to fix.

SupraLabs also tried fine-tuning the HiFi-GAN vocoder on the acoustic model's own predicted mel spectrograms instead of ground-truth ones, a common trick called GTA fine-tuning. On this release it made things worse, UTMOS dropped to 3.34 with noisier pitch, likely because LJSpeech is too small and too uniform a dataset for that approach to pay off here. The published model uses the ground-truth-mel vocoder only.

Training hardware

Both stages trained on a single RTX 5060 Ti with 16GB of VRAM. That is a training-time detail, not an inference requirement. Given the acoustic model and vocoder combine for under 44 million parameters, running inference on the same card, or a much smaller one, is not a stretch, though SupraLabs does not publish a separate inference benchmark.

Install it

The model card's own instructions call for Python 3.10 or newer, PyTorch with CUDA, and the Coqui-TTS package. The install pulls the inference script and model files individually from the Hugging Face repo with wget, rather than cloning a GitHub project.

FAQ

What is SupraTTS-0.1-Beta?

A small open-weight English text-to-speech model from SupraLabs, combining a 29.6 million parameter Glow-TTS acoustic model with a 14 million parameter HiFi-GAN vocoder. It is MIT licensed and trained on the LJSpeech dataset.

Can I run it locally right now?

Yes. The model card publishes a working install path: download the inference script and the two checkpoint files from the Hugging Face repo with wget, then run python3 infer_v2.py after installing Python 3.10+, PyTorch with CUDA, and the coqui-tts package.

How does it compare to Flare-TTS-v1.5?

SupraTTS-0.1-Beta scores 3.59 UTMOS against 3.05 for v1.5, and fixes a pacing bug that made v1.5 run about 8.7% faster than natural speech. Word error rate is about the same between the two.

Is it free to use commercially?

Yes. It ships under the MIT license, which permits commercial use, modification, and redistribution.

Does it support languages other than English?

No. It is trained only on LJSpeech, a single English speaker's voice, and the model card lists English as the only supported language.

Where to go from here

• Want an even smaller model known to run well on CPU? Run Kokoro-82M locally covers an 82 million parameter Apache-licensed alternative.
• Curious about the other TTS model that entered the trending top 20 around the same time? Fish Audio S1-mini overview covers a larger, non-commercial model with no local install path yet.
• Need multilingual voice cloning instead of a single English voice? Run IndicF5 locally covers an 11-language model from a 350 million parameter checkpoint.
• Compare SupraTTS-0.1-Beta against other open TTS models: browse the local AI model directory to see specs side by side.