TuniSpeech-21h

Summary

TuniSpeech-21h is a 21-hour speech corpus specifically designed for Tunisian Arabic (Derja). It was developed to address the underrepresentation of this dialect in the landscape of Automatic Speech Recognition (ASR). The dataset is compiled from social media (YouTube and Facebook) and broadcast materials, capturing a wide range of spontaneous speech and diverse linguistic characteristics.

Dataset Statistics

Feature	Count
Total Segments	32,294
Total Words	170,488
Total Unique Words	34,866
Average Utterance Length	5.281 words
Mean Clip Duration	7 seconds
Max Clip Duration	10 seconds

Creation Pipeline

The dataset was developed through a transparent and reproducible pipeline:

1. Data Collection: Content was sourced from YouTube and Facebook to ensure regional and demographic diversity.
2. Preprocessing: Includes audio normalization, noise reduction, and the removal of overlapping speech.
3. Segmentation: Automatic segmentation was performed based on natural pauses in speech, followed by manual verification.
4. Transcription: A semi-automatic approach was used: initial transcription via "TurboScribe" (Whisper-based) followed by manual review and correction using "Subtitle Edit".
5. Transcription Style: Clean Verbatim Transcription, which preserves the original message while omitting hesitations and filler words.

Baseline Benchmark

Several state-of-the-art ASR models were evaluated on this corpus. Whisper large-v2 currently holds the best overall performance on TuniSpeech-21h:

• Word Error Rate (WER): 24.74%
• Character Error Rate (CER): 8.32%
• Mixed Error Rate (MER): 16.53%

Citation

If you use this dataset in your research, please cite the following paper:

Sghaier, Mohamed Ali; Bellagha, Mohamed Lazhar and Zrigui, Mounir (2026). A New Tunisian Arabic Corpus and Benchmark for Automatic Speech Recognition. In Proceedings of the 18th International Conference on Agents and Artificial Intelligence.