README.md

---
license: apache-2.0
language:
- ar
tags:
- audio
- automatic-speech-recognition
---
<style>
img {
 display: inline;
}
</style>
![license](https://img.shields.io/badge/license-apache2-lightgrey)
|![Language](https://img.shields.io/badge/Language-Tunisian-lightgrey)
|[![Model architecture](https://img.shields.io/badge/Model_Arch-TDNN-lightgrey)](https://github.com/linagora-labs/ASR_train_kaldi_tunisian?tab=readme-ov-file#acoustic-model-am)
|[![GitHub](https://img.shields.io/badge/GitHub-ASRTrainKaldiTunisian-lightgrey)](https://github.com/linagora-labs/ASR_train_kaldi_tunisian)


# LinTO ASR Arabic Tunisia v0.1 

**LinTO ASR Arabic Tunisia v0.1** is an Automatic Speech Recognition (ASR) model for the Tunisian dialect,
with some capabilities of code-switching when some French or English words are used.

This repository includes two versions of the model and a Language model with ARPA format:
- `vosk-model`: The original, comprehensive model.
- `android-model`: A lighter version with a simplified graph, optimized for deployment on Android devices or Raspberry Pi applications.
- `lm_TN_CS.arpa.gz`: A language model trained using SRILM on a dataset containing 4.5 million lines of text collected from various sources.

## Model Overview

- **Model type**: Kaldi TDNN
- **Language(s)**: Tunisian Dialect
- **Use cases**: Automatic Speech Recognition (ASR)

### Model Performance

The following table summarizes the performance of the **LinTO ASR Arabic Tunisia v0.1**  model on various considered **test sets**:

| Dataset | CER     | WER     |
| :------- | :------- | :------- |
| [Youtube_TNScrapped_V1](https://huggingface.co/datasets/linagora/linto-dataset-audio-ar-tn#data-table) | `25.39%` | `37.51%` |
| [TunSwitchCS](https://huggingface.co/datasets/linagora/linto-dataset-audio-ar-tn#data-table) | `17.72%` | `20.51%` |
| [TunSwitchTO](https://huggingface.co/datasets/linagora/linto-dataset-audio-ar-tn#data-table) | `11.13%` | `22.54%` |
| [ApprendreLeTunisien](https://huggingface.co/datasets/linagora/linto-dataset-audio-ar-tn#data-table) | `11.81%` | `23.27%` |
| [TARIC](https://github.com/elyadata/TARIC-SLU) | `10.60%` | `16.06%` |
| [OneStory](https://huggingface.co/datasets/linagora/linto-dataset-audio-ar-tn#data-table)| `1.53%` | `4.47%` |

### Training code

The model was trained using the following GitHub repository: [ASR_train_kaldi_tunisian](https://github.com/linagora-labs/ASR_train_kaldi_tunisian)

### Training datasets

The model was trained using the following datasets:

- **[LinTO DataSet Audio for Arabic Tunisian](https://huggingface.co/datasets/linagora/linto-dataset-audio-ar-tn):** This dataset comprises a collection of Tunisian dialect audio recordings and their annotations for Speech-to-Text (STT) tasks. The data was collected from various sources, including Hugging Face, YouTube, and websites.
- **[LinTO DataSet Audio for Arabic Tunisian Augmented](https://huggingface.co/datasets/linagora/linto-dataset-audio-ar-tn-augmented):** This dataset is an augmented version of the LinTO DataSet Audio for Arabic Tunisian v0.1. The augmentation includes noise reduction and voice conversion.
- **[TARIC](https://github.com/elyadata/TARIC-SLU):** This dataset consists of Tunisian Arabic speech recordings collected from train stations in Tunisia.

## How to use

### 1. Download the model

You can download the model and its components directly from this repository using one of the following methods:

**Method 1: Direct Download via Browser**

1. **Visit the Repository**: Navigate to the [Hugging Face model page](https://huggingface.co/linagora/linto-asr-ar-tn-0.1).
2. **Download as Zip**: Click on the "Download" button or the "Code" button (often appearing as a dropdown). Select "Download ZIP" to get the entire repository as a zip file.

**Method 2: Using `curl` command**

You can follow the command below:

```bash
sudo apt-get install curl

curl -L https://huggingface.co/linagora/linto-asr-ar-tn-0.1/resolve/main/vosk-model.zip --output vosk-model.zip

```
(or same with `android-model.zip` instead of `vosk-model.zip`)

**Method 3: Cloning the Repository**

You can clone the repository and create a zip file of the contents if needed:

```bash
sudo apt-get install git-lfs
git lfs install

git clone https://huggingface.co/linagora/linto-asr-ar-tn-0.1.git

cd linto-asr-ar-tn-0.1
```

### 2. Unzip the model

This can be done in bash:
```bash
mkdir dir_for_zip_extract

unzip /path/to/model-name.zip -d dir_for_zip_extract
```

### 3. Python code

First, make sure to install the required dependencies:

```bash
pip install vosk
```

Then you can launch the inference script from this repository:
```bash
python inference.py <path/to/your/model> <path/to/your/audio/file.wav>
```

or use such a python code:
```python
from vosk import Model, KaldiRecognizer
import wave
import json

model_dir = "path/to/your/model"
audio_file = "path/to/your/audio/file.wav"

model = Model(model_dir)

with wave.open(audio_file, "rb") as wf:
    if wf.getnchannels() != 1 or wf.getsampwidth() != 2 or wf.getcomptype() != "NONE":
        raise ValueError("Audio file must be WAV format mono PCM.")
    
    rec = KaldiRecognizer(model, wf.getframerate())
    rec.AcceptWaveform(wf.readframes(wf.getnframes()))
    res = rec.FinalResult()
    transcript = json.loads(res)["text"]
print(f"Transcript: {transcript}")
```

## Example

Here is an example of the transcription capabilities of the model:

<audio controls>
  <source src="https://huggingface.co/linagora/linto-asr-ar-tn-0.1/resolve/main/sample.wav" type="audio/wav">
</audio>

### Result:
<p dir="rtl">
بالدعم هاذايا لي بثتهولو ال berd يعني أحنا حتى ال projet متاعو تقلب حتى sur le plan حتى فال management يا سيد نحنا في تسيير الشريكة يعني تبدل مية و ثمانين درجة ماللي يعني قبل ما تجيه ال berd و بعد ما جاتو ال berd برنامج نخصص لل les startup إسمو
</p>

## WebRTC Demonstartion

Install required dependencies:
```bash 
pip install vosk
pip install websockets
```

If not done, close the repostorory:
```bash
git clone https://huggingface.co/linagora/linto-asr-ar-tn-0.1.git
```

Then call the `app.py` script:
```bash
cd linto-asr-ar-tn-0.1/Demo-WebRTC

python3 app.py <model-path>
```
Access the web interface at: `localhost:8010` Just start and speak.

Preview of the web app interface:
![Demo Interface](https://huggingface.co/linagora/linto-asr-ar-tn-0.1/resolve/main/example.png)


## Citation

When using the LinTO ASR for Arabic Tunisian model, please cite the following paper (arxiv:2504.02604).
```bibtex
@misc{linagora2024Linto-tn,
  title = {LinTO Audio and Textual Datasets to Train and Evaluate Automatic Speech Recognition in Tunisian Arabic Dialect},
  author = {Hedi Naouara and Jérôme Louradour and Jean-Pierre Lorré},
  year = {2025},
  month = {March},
  eprint={2504.02604},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  note={Good Data Workshop, AAAI 2025},
  url={arxiv.org/abs/2504.02604},
}
```