You need to agree to share your contact information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Nava FA is a Persian (Farsi) text-to-speech student. Research and evaluation are free worldwide. Commercial use needs a written grant from Zorost Intelligence Company (contact@zorost.com) and is limited to products whose primary market is the United States, Canada, or Europe. Commercial use in any other country or region is prohibited.
Log in or Sign Up to review the conditions and access this model content.
Nava FA (zorost/nava-fa)
Persian / Farsi (fa, fas, فارسی) text-to-speech.
6-layer Pocket-TTS FlowLM student from Zorost Intelligence AI Lab.
CPU. Ten Lab characters. Optional 30-second voice clone.
This card is the Farsi TTS student you search for as Persian TTS, Farsi TTS, Persian speech, or تبدیل متن به گفتار فارسی.
- Article: https://zorost.com/ai-lab/fine-tuning/nava-fa
- Code and local app: https://github.com/zorost/nava-fa
- Demo:
nava-fa appon a laptop (Python 3.10+)
This is a Zorost student distilled from a Zorost 24-layer teacher trained on CC0 Persian audio. It is an honest beta.
Territorial commercial restriction
Research and evaluation are free worldwide.
Commercial use is limited to applications whose primary market and place of operation are in the United States, Canada, or Europe (EU, UK, EEA, Switzerland). Commercial use in any other country or region is prohibited. Zorost will not issue a grant for those markets.
A commercial product that is allowed still needs a written grant: contact@zorost.com. Hugging Face gating is access control, not checkout.
Full text: LICENSE-WEIGHTS.md (v1.1).
How to use
Request access, then:
pip install git+https://github.com/zorost/nava-fa.git
huggingface-cli login
nava-fa "دانشگاه صنعتی آریامهر" --character maryam -o aryamehr.wav
nava-fa app
from nava_fa import Nava
nava = Nava.from_pretrained("zorost/nava-fa")
wav, meta = nava.synthesize("دانشگاه صنعتی آریامهر")
open("aryamehr.wav", "wb").write(wav)
print(meta["rtf"])
wav, meta = nava.synthesize("سلام، حال شما چطور است؟", character="ali")
Clone from a clip you have the right to use. The model keeps the densest 30 seconds, not the first five:
wav, meta = nava.synthesize(
"دانشگاه صنعتی آریامهر",
voice_wav="my-voice.wav",
)
First run downloads about 418 MiB. Inference is CPU by default. You type Persian script. You do not phonemise by hand.
Characters
Ten Lab personas. Famous-feeling Persian first names. Adult CC0 Common
Voice speakers. Not celebrity likenesses. Default: forough.
| Id | Name | Role |
|---|---|---|
| forough | فروغ | Poetry |
| parvin | پروین | Story |
| leila | لیلا | Conversation |
| maryam | مریم | News |
| shirin | شیرین | Warm |
| ali | علی | Formal |
| reza | رضا | Everyday |
| amir | امیر | Narration |
| kian | کیان | Bright |
| bahram | بهرام | Warm |
nava-fa characters lists the cast. A user recording overrides the
picker for that request. Consent stays on you.
Model details
| Language | Persian / Farsi (fa) |
| Architecture | Pocket-TTS FlowLM, 6 layers, dim 1024, 16 heads |
| Codec | Kyutai Mimi, 24 kHz, 12.5 frames/s |
| Tokenizer | SentencePiece 4000, Nava transcripts |
| Size | about 418 MiB (model.safetensors) |
| Identity | student step 200000, eos -2 |
| SHA256 | 5908c607df7c5860b7de242370875fed741f0688583795cc815f80f2e071af42 |
| Front-end | zorost_fa (ezafe, numbers, yeh/kaf fold) |
The 24-layer teacher is a training artifact. It is not this repo.
Training data
CC0 Persian speech. Hub pointers, not a WAV dump:
- Mana-TTS
- GPTInformal-Persian
- Mozilla Common Voice Persian (filtered)
- Neyshekar
Do not rehost those audio trees. Dari and Tajik are not treated as this Persian student.
Evaluation (19 September 2026)
Same prompts, default voice, Whisper-large-v3 fa, house text norm.
Nava student step 200000 versus a CC BY-NC 6-layer Pocket-TTS Farsi
student (mehdi-hf/pocket-tts-farsi-v2), evaluation only:
| Suite | Nava corpus WER | comparator WER |
|---|---|---|
| gold-14 | 0.3165 | 0.4304 |
| dest50 | 0.2574 | 0.2684 |
dest50 mean CER 0.0879 versus 0.1048. Mean RTF dest50 0.574 versus 0.761.
This is not a studio-100 / 0.174 protocol and not a Common Voice 300 card. Those are different sets. Details in the GitHub docs/EVALUATION.md.
Honest beta. The Lab does not claim this is the best Farsi TTS on every published card.
Intended use
- Persian / Farsi speech for research
- Laptop and CPU demos
- Commercial products whose primary market is the United States, Canada, or Europe, after a written grant
Out of scope
- Commercial products whose primary market is outside the United States, Canada, and Europe
- Celebrity clones without a talent grant
- Treating Dari or Tajik as this Persian student
- Marking these weights MIT, CC BY-NC, or GPL
License stack
| Layer | Grant |
|---|---|
| Application code | MIT |
| These weights | Zorost Nava FA License v1.2: research free worldwide; commercial paid grant, United States, Canada, and Europe only |
Mimi tensors mimi.* |
CC BY 4.0, Kyutai Labs |
| Training audio | CC0 upstream. Do not rehost |
Commercial grant: contact@zorost.com.
Citation
See CITATION.cff. Also cite Kyutai Pocket-TTS and Mimi.
- Downloads last month
- -
Model tree for zorost/nava-fa
Base model
kyutai/pocket-tts