bongovaad is a Python package for transcribing Bengali audio from YouTube videos. It uses a fine-tuned Whisper model optimized for Bengali speech recognition.

• Fine-tuned Model: We've LoRA-tuned the whisper-large-v2 model on the 'bn' subset of Mozilla Common Voice 13, achieving a Word-Error-Rate(WER) of 57, compared to the WER of 103.4 in the original OpenAI paper (Page 23).
• SRT Subtitle Generation: Automatically creates SRT subtitle files for video players.
• Efficient Audio Processing: Handles audio segmentation for longer videos with progress tracking.
• Temporary File Management: Uses temporary directories for clean processing.
• Robust Error Handling: Comprehensive error handling and logging.
• Command-line Interface: Easy-to-use CLI with multiple options.

• Python 3.8 or higher
• CUDA-compatible GPU (recommended for faster processing)
• ffmpeg

Ubuntu/Debian:

sudo apt install ffmpeg -y

macOS (using Homebrew):

brew install ffmpeg

Windows (using Chocolatey):

choco install ffmpeg

pip install bongovaad

bongovaad --url "https://www.youtube.com/watch?v=VIDEO_ID"

bongovaad --url "https://www.youtube.com/watch?v=VIDEO_ID" \
          --segment-length 10 \
          --output-format both \
          --device cuda \
          --verbose

The tool generates two types of output files:

1. Text File (VIDEO_ID.txt): Contains the full transcription text.
2. SRT File (VIDEO_ID.srt): Contains time-coded subtitles compatible with video players.

You can also use bongovaad as a Python library:

from bongovaad import BongoVaadTranscriber

# Initialize the transcriber
transcriber = BongoVaadTranscriber(use_8bit=True, device="cuda")

# Transcribe a YouTube video
output_files = transcriber.transcribe(
    url="https://www.youtube.com/watch?v=VIDEO_ID",
    segment_length_seconds=10,
    output_format="both"
)

# Print output file paths
print(f"Text file: {output_files['txt']}")
print(f"SRT file: {output_files['srt']}")

• Processing time depends on video length and hardware capabilities.
• Using a GPU significantly improves performance.
• Longer segment lengths may improve speed but could reduce accuracy for complex audio.

This project is licensed under the MIT License. See the LICENSE file for details.

Contributions are welcome! Here's how you can contribute:

1. Fork the repository
2. Create a feature branch: git checkout -b feature-name
3. Commit your changes: git commit -am 'Add some feature'
4. Push to the branch: git push origin feature-name
5. Submit a pull request

• PEFT for parameter-efficient fine-tuning
• Whisper for the base ASR model
• Mozilla Common Voice for the training dataset