balbal-v2e-zellige

balbal-v2e-zellige

Turns Moroccan Darija speech into Arabic-script text with a timestamp for each word, on an ordinary CPU. An experimental preview for researchers and developers, under a non-commercial license.

Made by erayyapagci on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is a speech recognition model for Moroccan Darija. It turns audio into Arabic-script text with a timestamp on every word, and it runs on a CPU. The author calls it an experimental preview, and it is the first Arabic model in the balbal line.

Who it suits: It suits researchers and developers who need Darija transcription without a GPU. A non-commercial license applies, so it fits study and prototypes.

Title: Who runs it
Industry | Who uses it | How they use it | Example
Research | Speech researchers | Measure Darija accuracy against larger open models | Score a test set of Moroccan recordings
Software | Developers | Add Darija transcription to a tool that runs on a CPU | A script that transcribes a WAV file
Media | Caption editors | Get word timestamps to line text up with audio | Draft subtitles for a Moroccan interview

If you need a commercial product or another Arabic dialect, you can skip it.

Where to use it:

• Python 3.11 or 3.12 on Windows, macOS or Linux
• The balbal command line, version 0.6.0 or newer
• A CPU machine: the card reports about 100 times faster than real time on 8 threads
• A browser: the author links a demo page for trying it

Title: Jobs it does
Use case | Who | What they get back
Transcribing a recording | Researchers | Arabic-script text of the speech
Subtitling | Caption editors | The text with a timestamp for every word
Benchmarking | Machine learning teams | Error rates on three Moroccan Darija test sets

How to set it up:

• Python 3.11 or 3.12
• The balbal package, version 0.6.0 or newer
• No GPU is needed

1. Install the balbal package with pip.
2. Run it on a WAV file. The first run downloads this model.
3. To fetch the repository files by hand, install the Hugging Face Hub tools and download them.

pip install -U balbal
balbal run --model zellige speech.wav          # downloads this model on the first run

pip install -U huggingface_hub
hf download erayyapagci/balbal-v2e-zellige

How to use it: Run the balbal command on an audio file, or load the model from Python. Long recordings are split into 30 s pieces automatically. French and other Latin-script words are spelled with Arabic letters.

balbal run --model zellige --json speech.wav   # text + per-word timestamps

import balbal
model = balbal.load("balbal-v2e-zellige")
print(model.transcribe("speech.wav")["text"])

Title: Ways to run it
Tool | What it does
balbal run --model zellige | Transcribes an audio file to text
--json | Adds a timestamp for each word
balbal.load | Loads the model inside a Python script
transcribe | Returns the recognized text of a file

Example: You run the balbal command with the json flag on a short Darija recording named speech.wav. It prints the recognized speech in Arabic script, and each word carries its own timestamp.

For newcomers and experts: If you are new, speech recognition means turning spoken audio into written text. The first small thing to try is installing balbal and running it on one short WAV file of Darija.

If you know the field, the acoustic model is a FastConformer-CTC initialised from Sabeq-ASR. A two-layer looped refiner makes 3 passes over its output, and a product-key memory holds 193M of the 307M parameters. Decoding uses a 6-gram KenLM over SentencePiece pieces. The trade-off is accuracy: 42.85 WER on the Casablanca Morocco test set with the language model, 46.80 without it.

Audio is cut into 30 s chunks
FastConformer-CTC acoustic model
Looped refiner, 3 passes
Beam search with a 6-gram KenLM
Arabic text with word timestamps

Good to know:

Title: Average WER on three Moroccan test sets
balbal-v2e-zellige: 44.38%
balbal-v2e-zellige, no LM: 47.58%
Cohere Transcribe Arabic: 53.22%
Jais-ASR: 53.73%
MoulSot v0.3: 54.18%

• Preview: the author marks it experimental and asks for feedback.
• Tuning: the decoder settings were chosen using a single Moroccan validation set, Casablanca Morocco.
• Files: PyTorch weights are 1.2 GB and the ONNX graph is 1.25 GB.
• License: cc-by-nc-sa-4.0, so it is non-commercial.
• Cost to run: no API key, account or paid service is named.
• Data it reaches: the first run downloads the model, and the card names no other access.
• Not for: commercial use, which the license bars.

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Downloads in the last 30 days: 64
• Likes: 3
• License: cc-by-nc-sa-4.0
• Languages: ary, ar
• Based on: MostafaAhmed98/Sabeq-ASR
• Tags: onnx, speech-recognition, arabic, darija, moroccan-arabic, fastconformer, kenlm, automatic-speech-recognition, model-index

Links

• Model: https://huggingface.co/erayyapagci/balbal-v2e-zellige
• Files: https://huggingface.co/erayyapagci/balbal-v2e-zellige/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download erayyapagci/balbal-v2e-zellige

Then follow the usage in its model card. Nothing in this kit runs by itself.