BounharAbdelaziz/Moroccan-Darija-Youtube-Commons-Eval

Overview

This dataset is a curated subset of the evaluation dataset, filtered to include only audio samples with durations of less than 4 seconds. These shorter samples typically have high-quality transcriptions, making them ideal for evaluating speech-to-text (STT) models.

The primary motivation for creating this filtered dataset is to improve efficiency in model evaluation, as shorter audio clips require less computational overhead while still maintaining meaningful transcription tasks.

Dataset Summary

• Total Rows: 105
• Filter Applied: Removed audio samples with durations ≥4 seconds.

Dataset Structure

The dataset includes the following columns:

1. audio:

   • Contains the raw audio waveform and metadata.
   • Stored as a dictionary with:
     • array: A list or numpy array of raw audio samples.
     • sampling_rate: The sampling rate in Hz.

2. transcription:

   • The textual transcription corresponding to the audio clip.

3. language:

   • The language of the transcription.
   • Currently, this includes "moroccan_darija", with potential for expansion to include "amazigh" in the future.

4. dataset_source:

   • The name of the original dataset source for each sample.

5. duration:

   • The duration of the audio sample in seconds. This field is precomputed during the filtering process.

Example Row

{
  "audio": {
    "array": [0.001, 0.002, -0.001, ...],  // Raw waveform data
    "sampling_rate": 16000  // Sampling rate in Hz
  },
  "transcription": "و إلى شفتيها ها وجهي ها وجهك.",
  "language": "moroccan_darija",
  "dataset_source": "adiren7/darija_speech_to_text",
  "duration": 1.539563
}

Acknowledgments

This dataset is built on the foundation of the adiren7/darija_speech_to_text dataset (validation split to avoid data contamination). We extend our gratitude to the creators for providing such valuable data resources.