whatmodel.sh / tts / Nabra-82M v0.1
Nabra-82M v0.1
↗ share
+ contribute
Nabra-82M v0.1 is an open-weights text-to-speech model from oddadmix, released July 2026. Its published size is 0.082B parameters. Licence: apache-2.0.
Catalogue identity notes
oddadmix/Nabra-82M-v0.1 @ adf6abf35c46db5f2b08803b5067a12c759b1ee1 (created 2026-07-01, Apache-2.0): Kokoro-82M fine-tuned for Modern Standard Arabic, single female voice af_msa, needs diacritized input. Arena identity: Arabic TTS Arena model_id nabra_82m 'Nabra 82M'; the arena wrapper models/nabra_82m.py pins MODEL_REPO oddadmix/Nabra-82M-v0.1 at MODEL_REVISION adf6abf35c46db5f2b08803b5067a12c759b1ee1 (the same revision). Distinct from kokoro-82m (English base) and the MLX conversion oddadmix/Nabra-82M-MLX.
Model information sources
Release date ↗ · Weights licence ↗ · Parameter count ↗
Card: 'an 82M-parameter neural text-to-speech model' (Model Details: Parameters ~82M).
Score within the Vapi Humanness · Bradley–Terry rating, 2026-09-25 snapshot comparison set only.
$ how it compares
The text-to-speech models ranked either side of Nabra-82M v0.1, out of 148 with a score.
| # | model | score | where |
|---|---|---|---|
| 104 |  Parler-TTS Mini v1 | 44±6 | local |
| 105 |  Sonic (1) | 42±3 | comparison only |
| 106 | Nabra-82M v0.1 | 42±10 | local |
| 107 |  OpenAudio S1 Mini | 41±5 | api price unknown +local |
| 108 | Fish Speech 1.5 | 41±3 | local |
see all text-to-speech models · find the ones that fit your machine
$ what it's good at
other tests1 not in a task · 62% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Arabic TTS Arena · community blind A/B rating, 2026-09-24 snapshot | 965.0 ✎ | ▂▃··▃·█▅···▂ beats 1/12 | 62% |
$ where to run it
Other ways to run it
- Run it as published · needs ~0.33 GB · from huggingface.co/oddadmix (official)
pip install "kokoro @ git+https://github.com/Oddadmix/kokoro.git@main" torch soundfile huggingface_hub misaki camel-tools && camel_data -i disambig-mle-calima-msa-r13 && wget https://gist.githubusercontent.com/Oddadmix/dc699f7942a9516ce29d4842c7aed756/raw/827b541c892a862f9ef3b44006a6e27b100d1bdd/arabic_g2p.py && python -c "import numpy as np, torch, soundfile as sf; from huggingface_hub import hf_hub_download; from arabic_g2p import ArabicG2P, EXTRA_SYMBOLS, clean_phonemes, normalize_text; from kokoro import KModel, KPipeline; from kokoro import pipeline as kpipeline_mod; R = 'oddadmix/Nabra-82M-v0.1'; kmodel = KModel(repo_id=R, config=hf_hub_download(R, 'config.json'), model=hf_hub_download(R, 'kokoro_arabic.pth'), disable_complex=True).eval(); kmodel.vocab.update(EXTRA_SYMBOLS); kpipeline_mod.LANG_CODES.setdefault('ar', 'ar'); pipeline = KPipeline(lang_code='ar', repo_id=R, model=kmodel); _orig_g2p = pipeline.g2p; pipeline.g2p = lambda t: (clean_phonemes(_orig_g2p(t)[0]), _orig_g2p(t)[1]); voice = torch.load(hf_hub_download(R, 'af_msa.pt'), map_location='cpu', weights_only=True); text_norm, _ = normalize_text('مرحبا بك في نبرة'); diac = ArabicG2P(diacritize=True).diacritize(text_norm); sf.write('nabra.wav', np.concatenate([a.detach().cpu().numpy() for _, _, a in pipeline(diac, voice=voice, speed=1.0)]).astype(np.float32), 24000)"
Official weights and documentation: huggingface.co/oddadmix ↗
$ how the score was worked out · Arabic TTS Arena (Navid AI) ·… 62%
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
| source | results | share of the score |
|---|---|---|
 Arabic TTS Arena (Navid AI) · leaderboard snapshot 2026-09-24 | 1 | 62% |
| prior · pull toward the middle | 38% | |
| weighted average | = 41.5 |
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json