MohammadGholizadeh/youtube-farsi

📚 Unified Persian YouTube ASR Dataset (msghol/youtube-farsi)

This dataset is an enhanced and user-ready version of `PerSets/youtube-persian-asr`, restructured for seamless integration with Hugging Face Dataset Viewer and downstream ASR pipelines. It simplifies the data format by combining audio and transcription into unified records, removing the need for preprocessing scripts.

🔍 Overview

The dataset provides Persian-language audio-transcription pairs sourced from YouTube videos with Persian audio and subtitles. Each entry is formatted to support automatic speech recognition (ASR) tasks using tools like 🤗 Transformers, 🐸 SpeechBrain, 🐍 torchaudio, and others.

Unlike the original dataset, this version bundles audio and transcription together, allowing users to view, access, and load the data directly from the Hugging Face Hub without further preprocessing.

✅ Key Improvements

• —✅ Viewer-Friendly Format: Designed to work seamlessly with the Hugging Face Dataset Viewer interface.
• —✅ Unified Record Structure: Each sample includes both the audio data and its corresponding transcription in a single dictionary format.
• —✅ No Preprocessing Required: Eliminates the need for Python scripts to align or merge audio and subtitle data.
• —✅ Easier Loading: Ready for direct use in training and evaluation pipelines with datasets.load_dataset().

📦 Dataset Details

• —Language: Persian (Farsi)
• —Total Duration: ~245 hours
• —Number of Samples: 400,000+
• —Source: Persian YouTube content with subtitles
• —Genres: Wide range including casual vlogs, educational videos, interviews, discussions, and entertainment
• —Sampling Rate: 16 kHz
• —File Format: WAV (PCM)
• —Transcription: Accurate sentence-level text aligned with subtitle timestamps

📁 Dataset Structure

Each sample in the dataset has the following JSON structure:

{
  "audio": {
    "path": "path/to/audio.wav",
    "array": [...],
    "sampling_rate": 16000
  },
  "transcription": "متن مربوط به این بخش از فایل صوتی"
}