hananeek2/algerian-STT-dialect

Algerian Darija Speech-to-Text Dataset

Dataset Summary

A speech recognition dataset of spoken Algerian Darija (Algerian Arabic dialect), containing approximately 7,498 audio clips totaling around 9 hours and 42 minutes of transcribed speech. The dataset covers a wide range of naturalistic spoken language sourced from Algerian media including TV programs, social content, and interviews.

Dataset Details

Feature	Value
Language	Algerian Arabic (Darija / arq)
Task	Automatic Speech Recognition (ASR)
Audio Format	WAV, 16 kHz mono
Total Duration	9 hours 42 minutes (582 minutes)
Number of Samples	7,498
Transcription Language	Algerian Darija (Arabic script + code-switching)

Dataset Structure

Each example contains two fields:

• audio: A 16 kHz mono WAV audio file
• sentence: The corresponding transcription in Algerian Darija

{
    "audio": {
        "array": [...],          # float32 numpy array
        "sampling_rate": 16000
    },
    "sentence": "سنناقش من خلالها هذه التفاصيل"
}

Data Sources

The audio comes from multiple Algerian media sources, including:

• Algerian TV programs — talk shows, social programs, interviews (e.g. Aallametni El Hayat, Elli Fat Mat)
• Online video content — Algerian YouTubers and content creators
• Scripted/fictional content — Algerian dialect TV series (e.g. Achoura Al-Achor, Charika Taht Elmora9aba)

Usage

from datasets import load_dataset

ds = load_dataset("hananeek2/algerian-STT-dialect")

# Access a sample
sample = ds["train"][0]
print(sample["sentence"])
# Play audio: sample["audio"]["array"] at 16kHz

Training Example (with 🤗 Transformers)

This dataset is compatible with fine-tuning Whisper or wav2vec2 for Algerian Darija ASR:

from transformers import WhisperProcessor, WhisperForConditionalGeneration
from datasets import load_dataset

dataset = load_dataset("hananeek2/algerian-STT-dialect", split="train")
processor = WhisperProcessor.from_pretrained("openai/whisper-small")

Data Fields

Field	Type	Description
audio	Audio (16kHz)	Spoken audio segment
sentence	string	Transcription in Algerian Darija

Curation Notes

• Audio was segmented from longer recordings into short clips
• Each segment was aligned with its transcription
• Vocal isolation (voice/background separation via Demucs) was applied to improve audio quality
• Files were normalized to 16 kHz mono WAV

Citation

If you use this dataset, please credit the author:

@dataset{hananeek2_algerian_darija_stt,
  author    = {Hanane Elkorichi},
  title     = {Algerian Darija Speech-to-Text Dataset},
  year      = {2025},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/hananeek2/algerian-STT-dialect}
}

License

This dataset is released under CC BY 4.0. You are free to use, share, and adapt the data with appropriate credit.

Contact

• Email: hananemarouaelkorichi@gmail.com
• LinkedIn: www.linkedin.com/in/hanane-elkorichi-510999218