Audio Input
Result
Results will appear here...
Record Audio
Ready to record
Save to Dataset
YouTube Auto-Subtitle Collector
💡 100% Automatic Subtitle Extraction: Provide a YouTube URL. The system will automatically download the audio and extract the transcription/translation directly from YouTube's subtitle tracks. You can select different languages for transcription and translation!
How It Works
  • System extracts the video ID and queries YouTube's transcript API directly.
  • Fetches the subtitle file for your selected Transcription language.
  • Fetches the subtitle file for your selected Translation language.
  • Automatically strips timestamps and HTML tags to leave pure text.
  • Downloads audio via yt-dlp, converts to 16kHz WAV, and saves to dataset.
Note: If the video has no captions (manual or auto-generated), the system will reject it to ensure dataset quality.
Upload Dataset
💡 Whisper Data Best Practices:
  • Duration: 3 to 10 seconds per clip is optimal.
  • Quality: 16kHz sample rate, mono, minimal background noise.
  • Metadata: Adding a language column explicitly teaches the model the source language.
  • Auto-Split: The system automatically splits your data 80% Train, 10% Validation, 10% Test.
Registered Datasets
Test Dataset Baseline

Select a dataset and click "Test Selected" to see baseline transcription & translation performance before training.

Manual Training Configuration

Full Auto-Optimization

Automatically tests 3 configurations (Baseline, Regularized, Aggressive) to find the best hyperparameters and bring WER under 10%.

Idle. Click start to begin automated hyperparameter search.
Live Training Metrics
No active training
Trained Models

No custom models trained yet.

End-to-End Architecture & Methodology