The Trio Serie
Three small models that hear, understand and speak Haitian Creole, made to run on your own device: no internet, no subscription. Together they are Klara, a Kreyòl voice assistant. · thetrio.space
| model | what it does | |
|---|---|---|
| 👂 Tande · Ear | Oswald-ASR-0.6-V2 | Kreyòl speech to text, numbers written as digits. Qwen3-ASR 0.6B fine-tune, WER 9.1% on gold test sets, 90% of numbers right (V1: 10.3%, 71%) |
| 🧠 Konprann · Understand | makandal-multiple-v2 | short spoken answers in Kreyòl, French, Spanish, English, and tool calls. Gemma 3 1B fine-tune, right tool 93% |
| 🗣️ Pale · Talk | Fatiman-TTS-v1 | Kreyòl text to speech, nine voices from consenting, paid speakers. Pocket TTS fine-tune, reading error 2.3% |
Here Oswald and Makandal run on a shared GPU and Fatiman on the CPU, from the same ONNX files as the phone app.
Pale ak Klara ak mikwo a oswa ekri. Oswald tande w, Makandal reponn (li ka rele zouti), Fatiman li repons lan. · Speak into the microphone or type. Oswald hears you, Makandal answers (calling tools when it needs a fact), Fatiman reads the answer. Known limit: tool calls are reliable on a first request and less so later in a conversation, so start a Nouvo konvèsasyon for each new kind of question (what Klara remembers about you is kept). Fatiman reads every answer with a Kreyòl voice, French, Spanish and English included.
Pale oswa ekri an kreyòl, franse, panyòl oswa angle.
Klara poko sonje anyen sou ou. Di l non w oswa kote w rete.
Nothing remembered yet. Tell her your name or your town.
Memwa: di l non w ak kote w rete, peze Nouvo konvèsasyon, epi mande l tan an lakay ou. · Memory: tell her your name and town, press New conversation, then ask about the weather at home.
Anrejistre oswa voye yon odyo an kreyòl, jiska 2 minit. V2 ekri chif yo an chif (25 desanm 2025, 3:30, 50%). · Record or upload Kreyòl audio, up to 2 minutes. V2 (default) writes numbers as digits: dates, prices, times, phone numbers, percentages; V1 writes them in words, pick it to compare. The examples are held-out clips of human read speech from CMU (licence in examples/LICENSE-CMU.txt). Oswald was not trained to recognize French or English.
Ekri an kreyòl, chwazi yon vwa. Chif yo li an kreyòl (25% → vennsenk pousan). · Type Kreyòl text and pick a voice. Numbers, percentages, times and units are spelled out in Kreyòl first. Up to about 1,200 characters. The nine voices belong to speakers who signed their consent.
| Tèks · Text | Vwa · Voice |
|---|
The same question, no tools, to makandal-multiple-v2 (left) and the original gemma-3-1b-it (right).