whisper-tunisian-dialect

Whisper Tunisian Dialect ASR (TuniSpeech‑21h)

Whisper Tunisian Dialect ASR

Overview

Whisper Tunisian Dialect ASR is a fine-tuned speech recognition model based on OpenAI Whisper-large-v2, adapted to the Tunisian Arabic dialect using the TuniSpeech corpus (~21 hours).

The model is designed for:

• Automatic Speech Recognition (ASR) in Tunisian dialect
• Transcription of short spontaneous speech
• Research on low-resource Arabic dialects

Repository:
https://huggingface.co/TuniSpeech-AI/whisper-tunisian-dialect

---

Model Architecture

• Base model: Whisper-large-v2
• Fine-tuning strategy: LoRA → merged into full model weights
• Framework: Hugging Face Transformers
• Decoding:
  • Forced language: Arabic (ar)
  • Task: transcription (not translation)

After training, LoRA adapters were fully merged, producing a standalone Whisper checkpoint compatible with standard inference pipelines.

---

Training Data

TuniSpeech Corpus

• Language: Tunisian Arabic dialect
• Duration: ~21 hours
• Recording conditions:
  • Real speech\
  • Spontaneous pronunciation\
  • Speaker variability

The dataset targets dialectal phonetics and vocabulary, which are poorly covered by standard Arabic ASR systems.

---

Experimental Setup

• Input audio:
  • Resampled to 16 kHz mono
• Training method:
  • Parameter-efficient fine-tuning (LoRA)
  • Followed by weight merging

Evaluation focused on:

• Dialectal word recognition
• Short utterance transcription
• Real-world speech robustness

---

Usage

Installation

pip install torch transformers librosa gradio

Python Inference

from transformers import pipeline

pipe = pipeline(
    "automatic-speech-recognition",
    model="TuniSpeech-AI/whisper-tunisian-dialect"
)

result = pipe("audio.wav")
print(result["text"])

---

Hugging Face Space

An interactive demo is available:

Upload or record Tunisian speech → get transcription instantly.

⚠️ On CPU, processing is slow.
Recommended:

• Use short audio (< 5 s)

---

Citation

If you use this model, please cite the associated work on:

Whisper fine-tuning for Tunisian dialect speech recognition using the TuniSpeech-21H corpus.

@inproceedings{tunispeech_whisper_2026,
  title     = {A New Tunisian Arabic Corpus and Benchmark for Automatic Speech Recognition},
  author    = {Sghaier, Mohamed Ali and Bellagha, Mohamed Lazhar and Zrigui, Mounir},
  booktitle = {In Proceedings of the 18th International Conference on Agents and Artificial Intelligence},
  year      = {2026}
}