shreeshacharya/Dhravani

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

Dataset Preparation Interface for Fine-tuning Whisper

A web-based interface for preparing audio datasets to fine-tune OpenAI's Whisper model. This tool helps in recording, managing, and organizing voice recordings with their corresponding transcriptions, with support for cloud storage and authentication.

Features

• 🔐 User authentication via Pocketbase
• ☁️ Cloud storage support (Hugging Face Datasets)
• 🌐 Multi-language support with native names
• 🎤 Modern Material Design recording interface
• 📝 CSV transcript file support
• 🎯 Session-based recording workflow
• 🔄 Advanced recording controls
• ⌨️ Keyboard shortcuts for efficiency
• 📊 Progress tracking and navigation
• 💾 Local and cloud metadata management
• 🎨 Responsive, mobile-friendly UI

Getting Started

1. Create a transcript CSV file with your content:

transcript
"First sentence to record"
"Second sentence to record"
# For multi-language support:
transcript_en,transcript_es
"English sentence","Spanish sentence"

2. Start the Flask application:

python app.py

3. Access the interface:

http://localhost:5000

Usage

1. Authentication

   • Sign in using your Google account

2. Session Setup

   • Upload your transcript CSV
   • Select language and recording location
   • Enter speaker details
   • Click "Start Session"

3. Recording

   • Use on-screen controls or keyboard shortcuts:
     • R: Start recording / Stop recording
     • Space: Play recording
     • Enter: Save recording
     • Backspace: Re-record
     • ←: Previous transcript
     • →: Skip current
   • Navigate using row numbers
   • Adjust transcript font size as needed

Data Storage

Recordings are stored in language-specific directories:

• Storage:

  datasets/
  ├── en/
  │   ├── audio/
  │   │   ├── {user_prefix}_{YYYYMMDD_HHMMSS}.wav
  │   │   └── ...
  │   └── en.parquet         # English recordings metadata
  ├── es/
  │   ├── audio/
  │   │   ├── {user_prefix}_{YYYYMMDD_HHMMSS}.wav
  │   │   └── ...
  │   └── es.parquet         # Spanish recordings metadata
  │
  

Technical Details

Audio Recording

• Browser Recording Format: 48kHz mono WebM
• Storage Format: 16bit mono WAV
• Maximum Duration: 30 seconds
• Audio Processing: WebM -> WAV conversion with sample rate adjustment
• Channels: 1 (mono)

Data Management

• Metadata Organization:
  • stats.json: Global recording statistics
  • {language_code}.parquet: Language-specific metadata files
• File Naming: {user_id_prefix}_{YYYYMMDD_HHMMSS}.{format}
• Unicode Handling: NFC normalization for text

Authentication

• Provider: Pocketbase with Google OAuth
• Session Management: Server-side Flask sessions

Languages

• Support: 74 languages with native names
• Codes: ISO 639-1 standard
• CSV Format:
  • Single language: transcript column
  • Multi-language: transcript_${lang_code} columns

Upload Management

• Queue System: Background worker thread
• Status Tracking: Real-time upload status polling
• Error Handling: Automatic retries with timeout
• Progress Updates: Toast notifications
• Temporary Storage: ./temp folder for conversions

Frontend Features

• Keyboard Shortcuts: Recording and navigation
• Real-time Status: Progress tracking and notifications

Security

• Authentication Required: All routes except static/login
• File Validation: MIME type and extension checking
• Secure Context: HTTPS recommended

Performance

• Upload Queue: Asynchronous processing
• Audio Conversion: Server-side processing
• Session Caching: Browser storage optimization
• Progress Tracking: Real-time websocket updates

Browser Support

• Chrome (recommended)
• Brave
• Edge
• Safari

Known Limitations

• Requires microphone permissions
• Internet connection needed
• Maximum recording duration: 30 seconds
• File size limits based on storage backend