All stories
SupraTTS-1.0-Preview Launches Community reports

SupraTTS packs an English voice into a 1M-parameter core

SupraLabs' downloadable speech preview reports roughly 85 times real-time CPU synthesis, with small automatic-scoring tests and an English-only scope.

By AITrending2 min read
Sources (4)

Why now The public preview surfaced in the October 17 discovery refresh. Its Hub history begins October 15, and an October 17 documentation update explains the current claims; neither timestamp establishes the exact public launch time.
SupraTTS author's three-panel chart plots automatic speech scores against parameter counts, with separate markers for comparisons using different evaluation sets.
Author-published automated-scoring comparison, not an AITrending reproduction or human listening test. Full chart and protocol notes preserved. Some rows use different evaluation sets; the chart's shared-front-end footnote also differs from the current card's bundled-front-end description, so it is not a verified comparison of total installed system sizes. Source: SupraLabs / arthurblg1802

A speech model aimed at ordinary CPUs

SupraLabs is pitching SupraTTS-1.0-Preview as an unusually small English text-to-speech system: its three-stage core has 1,031,754 parameters, according to the released configuration. The downloadable preview produces 24 kHz mono audio and ships a Python runtime, weights, sample recordings and benchmark scripts on Hugging Face.

The author reports a median real-time factor of 0.0118 on a Ryzen 5 3500X in a 24-sentence test, equivalent to roughly 85 seconds of generated speech per second of computation. That is a specific author-run CPU measurement, not a guarantee for every machine or sentence. The card compares it with the earlier SupraTTS-0.1-Beta on the same CPU and reports approximately 44 times faster synthesis.

The three stages predict phoneme durations, build a mel spectrogram and decode that representation into a waveform. A separate bundled English grapheme-to-phoneme front end uses dictionaries and a small fallback for unfamiliar words. The card describes runtime and weights as MIT-licensed, with Apache-2.0 dictionaries. It does not present this preview as a multilingual voice-cloning service.

The headline quality claim needs its test conditions

The model card calls the system “The new quality-per-parameter Pareto frontier for open TTS.” Its evidence is automatic scoring, including SCOREQ, UTMOS and DNSMOS, on a small 24-sentence evaluation. The table places the preview below the larger sanoTTS amy model on SCOREQ and UTMOS, while above it on DNSMOS-SIG. This supports a reported size-versus-score tradeoff within that comparison, not a universal listening-quality victory.

The card also says its predecessor’s quality results use different sentences and seeds, so those quality rows are directional rather than a controlled head-to-head result. The author-published chart marks some additional comparison rows as using different evaluation sets. These are predicted quality scores, not ratings from a new human listening panel. AITrending has not run the checkpoint or independently reproduced the speed or quality numbers.

Download, then synthesize locally

The documented route uses hf download SupraLabs/SupraTTS-1.0-Preview --local-dir supratts-1-0-preview, followed by installing the repository’s requirements. Those list NumPy, PyTorch 2.0 or newer, and SoundFile. The guide then loads MicroTTS from the downloaded directory, synthesizes a string and saves a WAV file. The author says inference needs no network access after setup.

The repository was created October 15 and includes an initial-release commit explicitly labelled private. An October 17 documentation edit brought it into this discovery pass; that edit should not be mistaken for a newly trained checkpoint or a known first-publication date. A larger version and additional languages remain roadmap items, not capabilities shipped in this English preview.

Related reading