erayy / lab

Experimental preview

balbal-v2.1-zellige

Open Moroccan Darija speech recognition that runs on a CPU. Speak Darija, get Darija text.

307M parametersArabic script outputCC BY-NC-SA 4.0

Unlike the other balbal demos, this page saves what you record or upload and uses it to train and test future balbal speech recognition models for Moroccan Darija. Models trained with it may be released publicly; the recordings themselves are never published.

  • Saved: the audio (up to 30 seconds), its transcript and any correction you send, the time, and a random ID of this browser.
  • Not saved: your name, e-mail or IP address.
  • Your choice: every transcription comes with a recording ID and a delete button. To remove a recording later, e-mail its ID to hello@erayy.com.

Vos enregistrements sont conservés. Tout ce que vous enregistrez ou téléversez ici (l’audio, jusqu’à 30 secondes, et sa transcription) est conservé et utilisé pour entraîner et améliorer les modèles balbal. Les enregistrements eux-mêmes ne sont jamais publiés. Nous ne conservons ni votre nom, ni votre e-mail, ni votre adresse IP. Vous pouvez supprimer un enregistrement grâce à son identifiant.

تسجيلاتك تُحفَظ. كل ما تسجّله أو ترفعه هنا (الصوت حتى 30 ثانية ونصّه المكتوب) يُحفَظ ويُستخدَم لتدريب نماذج balbal وتحسينها، ولن ننشر التسجيلات نفسها أبداً. لا نحفظ اسمك ولا بريدك الإلكتروني ولا عنوان IP الخاص بك. يمكنك حذف أي تسجيل بواسطة رقمه.

1 Record

Connecting…

Tick the consent box to start.

Try: واش نتا مزيان؟ · بغيت نمشي للسوق غدا الصباح

Up to 30 seconds · WAV, MP3, M4A, OGG or WebM

2 Transcript

The Darija transcript will appear here.

How it compares

Word / character error rate (%), lower is better. Full test sets: Casablanca MA test (1,045 clips), MGB-5 test (5,484), GigaSpeechBench MAR (15,961).

Model Params Average Casablanca MGB-5 GigaSpeechBench MAR
WERCERWERCERWERCERWERCER
balbal-v2.1-zellige307M44.3817.7442.8514.1145.8320.9644.4518.15
↳ without LM307M47.5819.0446.8015.6148.7622.1947.1719.31
Cohere Transcribe Arabic2B53.2221.5253.8617.4947.8721.0357.9326.04
Jais-ASR6B53.7321.9953.6219.7956.9325.4050.6320.79
MoulSot v0.32B54.1824.4949.8718.9261.4131.8451.2722.70

Run it yourself

pip install balbal
balbal run --model zellige speech.wav

Offline on a CPU (Linux, macOS, Windows). Weights on Hugging Face ↗, code on GitHub ↗. The model is released under CC BY-NC-SA 4.0 (non-commercial). Fine-tuned from Sabeq-ASR ↗.