Introduction

A compact Haitian Creole (Kreyòl) voice just landed for phones and browsers. Fatiman-mini-V1, from Hugging Face developer jsbeaudry, fine-tunes Gradium’s open phonon-31m-en flow language model for Kreyòl and keeps Kyutai’s Pocket TTS Mimi codec. Weights and voices went up on Hugging Face early on November 3, 2026 (UTC).

Primary artifact: jsbeaudry/Fatiman-mini-V1. It is the small sibling of Fatiman-TTS-v2 — taught by that larger system, more intelligible on the author’s voice-lab sweep, as close to the speaker, and a notch less natural.

This sits next to ArtRealmAI’s earlier Gradium Phonon English dispatch: same on-device family, new language lane.

What Shipped

  • Model: Fatiman-mini-V1 by jsbeaudry — Gradium Phonon TTS for Haitian Creole (ht).
  • Base: gradium/phonon-31m-en (6×512 flow LM), with Pocket TTS’s Mimi codec instead of Gradium’s.
  • Size: about 41.1M params; model.safetensors ~103 MB f32; phone/browser model.q8.gguf ~65 MB.
  • Audio: 24 kHz mono via Mimi (12.5 Hz frames).
  • Voices: eleven fixed embeddings in voices/ — seven consented Fatiman studio speakers (female1–3, female5, man1–3) plus four synthetic nouvo1–4. No audio-prompt voice cloning.
  • Runtimes: Gradium’s ptts stack — Rust CLI, Python, Android, iOS, and browser WebAssembly (phonon-tts).
  • Speed (author, Apple M3 Pro): q8 about 22.5× realtime on CPU through ptts; about 13.6× inside Chrome via WASM with threads on. Card headline: ~20× laptop CPU, 10–15× in-browser.
  • License: CC BY 4.0 (credit Pocket TTS by Kyutai). Phonon-31m-en is Apache 2.0 underneath.

Quality Numbers (Author Sweep)

On the same 12 assistant-style sentences, seeds, and five studio speakers as Fatiman’s production sweep (60 takes each; CER by Oswald V3; naturalness = UTMOS; similarity vs each recording’s own ceiling):

  • Fatiman-mini-V1 (t 0.3, 1 flow step, q8): score 0.728, UTMOS 2.59, similarity 0.817, CER 2.87%, 0/60 cut short
  • Fatiman-TTS-v2 (t 0.3, 4 steps): score 0.761, UTMOS 3.11, similarity 0.805, CER 5.16%

Paired takeaway from the card: mini is more intelligible, about as close to the speaker, and less natural — that naturalness gap is most of the remaining score difference. A wider 1,757-take eval at temperature 0.3 reports score 0.752, CER 2.99%, UTMOS 2.90, similarity 0.809.

These are author-reported numbers. Listen before you ship.

How to Run It

pip install ptts huggingface_hub
hf download jsbeaudry/Fatiman-mini-V1 --local-dir Fatiman-mini-V1

import sys
import ptts

sys.path.insert(0, "Fatiman-mini-V1")
from kreyol_text import normalize

tts = ptts.TTS("Fatiman-mini-V1", lang="none", quant="q8")
tts.save(
    "bonjou.wav",
    normalize("Bonjou! Li fè 31 degre jodi a."),
    voice="female1",
    temperature=0.3,
)

Always run text through normalize and keep lang="none" — the model was trained on words, not digits. Temperature 0.3 and one flow step are the production settings. ptts 0.5+ reads config.json fields such as time_rms_norm: "rms" and eos_threshold: -2.0.

Browser path

With phonon-tts, load hf://jsbeaudry/Fatiman-mini-V1 at quant: 'q8' and synthesize with a named voice. Write numbers out as Kreyòl words on the page too — the Rust normalizer has no Kreyòl path yet.

Pro Tip

Prefer female1, female2, or man1 first — the card says they score best (~0.81) on the wide eval. Keep consent: the seven real Fatiman voices are published with signed permission; the card asks you never to use a voice to deceive.

Where It Fits

If you need offline Kreyòl speech on a phone, laptop CPU, or a browser tab without an API key, Fatiman-mini is the clearest same-day open path: small q8 file, fixed voices, measured CER, CC BY 4.0. It will not clone from a reference clip, and naturalness trails the larger Fatiman-TTS-v2.

For other tiny open voices this week, see PiVo-1.4M, TiHon-TTS, and the English Phonon base above. For whole-dialogue English/German podcast TTS, see this morning’s Humaneness Voice Large NotebookLM Edition.

Conclusion

A 65 MB q8 Haitian Creole Phonon that reads ~20× realtime on a laptop CPU — and still runs in the browser — is a real on-device language drop, not a thin port. Pull the card, normalize a short Kreyòl line, and hear which of the eleven voices fits your assistant.

—Anabel ♡