phoonnx-f5tts
ONNX exports of F5-TTS / Habibi-TTS flow-matching voices for use with
phoonnx's F5TTSAdapter
(engine: "f5tts").
Each voice is exported with DakeQQ/F5-TTS-ONNX
into three ONNX graphs (model.onnx = transformer, F5_Preprocess.onnx,
F5_Decode.onnx) plus a config.json and the model's vocab.txt.
Voices
| Directory | Source model | Language | License |
|---|---|---|---|
f5tts-v1-base/ | SWivid/F5-TTS — F5TTS_v1_Base | multilingual (EN, ZH, +) | CC-BY-NC-4.0 — non-commercial only |
habibi-tts-unified/ | SWivid/Habibi-TTS — Unified | Arabic (all dialects, recommended) | CC-BY-NC-SA-4.0 — non-commercial, share-alike |
habibi-tts-msa/ | Habibi-TTS — Specialized/MSA | Arabic (Modern Standard) | Apache-2.0 |
habibi-tts-egy/ | Habibi-TTS — Specialized/EGY | Arabic (Egyptian) | Apache-2.0 |
habibi-tts-sau/ | Habibi-TTS — Specialized/SAU | Arabic (Saudi) | CC-BY-NC-SA-4.0 |
habibi-tts-uae/ | Habibi-TTS — Specialized/UAE | Arabic (Emirati) | CC-BY-NC-SA-4.0 |
habibi-tts-alg/ | Habibi-TTS — Specialized/ALG | Arabic (Algerian) | Apache-2.0 |
habibi-tts-irq/ | Habibi-TTS — Specialized/IRQ | Arabic (Iraqi) | Apache-2.0 |
habibi-tts-mar/ | Habibi-TTS — Specialized/MAR | Arabic (Moroccan) | Apache-2.0 |
namaa-saudi-tts-v2/ | NAMAA-Space/NAMAA-Saudi-TTS-V2 — Habibi-TTS fine-tune | Arabic (Saudi/Najdi) | CC-BY-NC-SA-4.0 — non-commercial, share-alike |
silma-tts-v1/ | silma-ai/silma-tts | Arabic (MSA, full tashkeel) + English | Apache-2.0 |
Licensing follows the upstream model cards: f5tts-v1-base is CC-BY-NC-4.0
(attribution: SWivid/F5-TTS); per
SWivid/Habibi-TTS (attribution:
SWivid/X-LANCE), Unified, SAU and UAE are CC-BY-NC-SA-4.0 (SADA/Mixat
restriction) while ALG, EGY, IRQ, MAR and MSA are Apache-2.0;
namaa-saudi-tts-v2 is a Habibi-TTS fine-tune by NAMAA-Space and inherits
the CC-BY-NC-SA-4.0 restriction from its Habibi/SADA/Mixat base; silma-tts-v1
is an independent pretrain by SILMA AI and is Apache-2.0 (commercial use
OK). See each voice's own README.md.
This repo's top-level license: tag (CC-BY-NC-4.0) is the conservative
single value required by the Hub's metadata field — it reflects the most
restrictive checkpoint. Always check the per-checkpoint table above before
using a specific voice commercially; several checkpoints (MSA, EGY, ALG,
IRQ, MAR, silma-tts-v1) are Apache-2.0 and do not carry the NC restriction.
The Unified voice supports dialect control via the dialect engine param
(UNK/MSA/SAU/UAE/ALG/IRQ/EGY/MAR) — see its README.
Listening samples
The samples/
folder has 2 wavs per voice, all synthesized through the phoonnx adapter, with
ASR round-trip intelligibility scores in
samples/RESULTS.md
(faster-whisper small; all 18 samples pass, CER ≤ 0.16).
Usage
All voices are in-context cloning models — synthesis requires a short
reference clip and its transcription alongside the text to speak. See
phoonnx's F5-TTS docs
for a full example (library-level and the ovos-tts-plugin-phoonnx config).