به نظر می‌رسد که بزها ابتدا حدود ده هزار سال پیش در کوه‌های زاگرس ایران اهلی شده باشند
Model outputMatches the reference
New: an open streaming Persian speech recognition model →
Persian speech AI. Open, accurate, real-time.
Speech recognition and text-to-speech models for Persian (Farsi), trained by Oncilla AI. Open weights, 8.8% word error rate on FLEURS, and it runs on a laptop.
Built on NVIDIA Nemotron and NeMo. Trained on 1,181 hours of openly licensed Persian speech.
- FLEURS word error rate
- 8.8%
- Training speech
- 1,181 h
- Lowest latency
- 0.3 s
- Speed on a laptop
- 10× real time
- Weights
- Open
Speech to text
Listen, then read what the model heard.
Four clips from the FLEURS test set, speech the model never trained on. Press play, then compare with the transcript below it.
اینترنت عناصر ارتباط جمعی و بین فردی را با هم ترکیب می‌کند
Model outputMatches the reference
کاری غذایی است که با گیاهان و ادویه‌ها همراه با گوشت یا سبزیجات تهیه می‌شود
Model outputMatches the reference
گیاهان با فتوسنتز غذای خود را از خورشید تهیه می‌کنند آن‌ها همچنین سایه فراهم می‌کنند
Model outputMatches the reference
These are four of the clips the model gets exactly right, picked to show what it does at its best. Across the whole
FLEURS test set it makes 8.8 word errors per 100 words (see the results below). “Matches the reference” is after
standard Persian text normalization, which ignores punctuation and the half-space (ZWNJ) versus space.
Audio:
FLEURS Persian test set, © Google, CC BY 4.0.
Live
Transcribes as you speak
Streaming recognition with 0.3 to 1.1 seconds of delay. Words appear while the speaker is still talking.
Files
Any audio or video file
Podcasts, interviews, films. Long recordings are processed in a stream, so length is not a problem.
Anywhere
Runs on a laptop
About 10 times faster than real time on an Apple M1, on CPU or GPU, or on your own servers. No data leaves your machine.
Text to speech
Natural Persian voices from text.
A 109.5M-parameter Persian speech synthesis model with voice cloning from a few seconds of audio. It runs on a CPU.
سلام، حال شما چطور است؟ امیدوارم روز خوبی داشته باشید.
Greeting · female voiceCloned voice
مادر کتاب را روی میز اتاق گذاشت و پنجره را باز کرد.
One sentence · female voiceCloned voice
شرکت آنتروپیک مدلی اقتصادی منتشر کرده که نشان می‌دهد هوش مصنوعی چگونه می‌تواند اقتصاد آمریکا را تا سال ۲۰۳۰ تغییر دهد.
News sentence · male voiceCloned voice
- Model size
- 109.5M parameters
- Trained on
- 973 h, 2,978 speakers
- Voice cloning
- From a short prompt
- Runs on
- A plain CPU
Each voice is cloned from a five-second prompt of a Mozilla Common Voice speaker (CC0) that the model never saw in training. Press play on the prompt, then on the generated speech. The open weights are released under CC BY-NC 4.0 (non-commercial use).
Results
The most accurate open Persian speech recognition we know of.
Word error rate on the FLEURS Persian test set, where lower is better. Within about a point of the best commercial systems, with weights you can run yourself.
| Model | Access | FLEURS Persian | Conversational speech |
|---|---|---|---|
| ElevenLabs Scribe v1Commercial API | Closed | 7.8% | not reported |
| Gemini Flash 2Commercial API | Closed | 8.0% | not reported |
| Oncilla AI · nemotron-asr-streaming-farsiStreaming, 600M parameters | Open weights | 8.8% | 26.0% |
| Shenava Koochik v1.0Streaming, 114M parameters | Open weights | 10.6% | not reported |
| Whisper large-v3 | Open weights | 24.0% | not reported |
| NVIDIA stt_fa FastConformerPrevious best NVIDIA Persian model | Open weights | 24.2% | 54.4% |
Our model and NVIDIA's were scored by us on the same 852 FLEURS Persian test clips with the same Persian text normalization. The other numbers are the ones their publishers report (ElevenLabs, Gemini Flash 2 and Whisper large-v3: elevenlabs.io/speech-to-text/persian; Shenava: its Hugging Face model card), and each used its own text normalization, so small differences are not meaningful. “Conversational speech” is a held-out test set of 9,154 YouTube and film clips (about 8 hours) that none of the models trained on.
What it is for
Where Persian speech technology helps.
Subtitles and captions
Turn Persian video and audio into timed text for broadcasters, creators and educators.
Transcription and search
Make interviews, podcasts and archives searchable, from newsroom back catalogs to oral histories.
Voice-overs and audiobooks
Generate narration in Persian without booking a studio, including cloned voices.
Accessibility
Live captions for deaf and hard-of-hearing people, and spoken versions of text for blind readers.
Calls and meetings
Transcribe customer calls and meetings in Persian for quality review and analytics.
Language learning
Pronunciation practice and listening exercises with accurate transcripts.
Open models
Everything is open. Take it, run it, check it.
Speech recognition
The model, the full training recipe, evaluation scripts and a local web app for live microphone transcription.
Weights: OpenMDW-1.1 · Code: Apache-2.0Speech synthesis
A compact Persian text-to-speech model with voice cloning, and the code that trained it.
Weights: CC BY-NC 4.0 · Code: MITBuilding with Persian speech?
The open models are free to use under their licenses. If you need them adapted to your audio, vocabulary or accents, or running privately on your own infrastructure, write to us.
mehdi@oncillaai.com →
Built by Oncilla AI
Mehdi Allahyari
Co-founder, Oncilla AI · AI researcher & system architect
Mehdi trained and released these models at Oncilla AI. He has spent 15+ years building machine-learning systems in production, holds a PhD in Computer Science, and was a tenure-track Assistant Professor researching NLP.