Voice Data Collection App

A web application for collecting voice recordings to fine-tune Whisper speech recognition models. Extract random phrases from PDFs or pasted text, record yourself speaking them, and automatically save WAV files with a CSV manifest.

Features

• Dual Input: Upload PDF files or paste text directly
• Smart Phrase Extraction: Automatically extracts random 3-5 word phrases
• Session-Based: Each recording session gets its own folder
• Auto-Save: Recordings are saved immediately after each phrase
• WAV Format: High-quality uncompressed audio for ML training
• CSV Manifest: Automatic mapping of phrases to audio files
• Keyboard Shortcuts: Space to record/stop, Enter for next phrase
• Progress Tracking: Visual progress bar and history panel

Installation

# Clone or navigate to the project
cd VoiceApp

# Install dependencies with uv
uv sync

Usage

Start the Server

uv run uvicorn main:app --port 8000

Or run directly:

uv run python main.py

Record Voice Data

1. Open http://localhost:8000 in your browser
2. Enter a session name (e.g., my_voice_session)
3. Choose input source:
   • Upload PDF: Drag & drop or click to select a PDF file
   • Paste Text: Switch to the text tab and paste your content
4. Set the number of phrases to extract (default: 50)
5. Click "Extract Phrases" to begin
6. For each phrase:
   • Click Record (or press Space) to start recording
   • Speak the phrase clearly
   • Click Stop (or press Space) to stop
   • Preview the recording if needed
   • Click Next Phrase (or press Enter) to save and continue
   • Use Re-record if you want to try again
7. Click Stop Session when finished (or complete all phrases)

Output

Each session creates a folder in data/<session_name>/ containing:

data/my_voice_session/
├── 001.wav
├── 002.wav
├── 003.wav
├── ...
└── manifest.csv

The manifest.csv file contains:

phrase,audio_file
"the quick brown",001.wav
"fox jumps over",002.wav
"the lazy dog",003.wav

Project Structure

VoiceApp/
├── app/
│   ├── __init__.py
│   ├── pdf_processor.py  # PDF text extraction & phrase generation
│   └── session.py        # Session management & file saving
├── static/
│   ├── index.html        # Main UI
│   ├── css/
│   │   └── styles.css    # Styling
│   └── js/
│       ├── app.js        # Main application logic
│       ├── recorder.js   # Audio recording module
│       └── wav-encoder.js # WAV format encoder
├── data/                 # Recording sessions (created at runtime)
├── main.py               # FastAPI server
├── pyproject.toml        # Project dependencies
└── README.md

API Endpoints

Method	Endpoint	Description
GET	/	Serve the web UI
GET	/api/health	Health check
POST	/api/start-session	Start a new recording session
POST	/api/upload-pdf	Upload PDF and extract phrases
POST	/api/parse-text	Parse pasted text and extract phrases
POST	/api/save-recording	Save a WAV recording
POST	/api/end-session	End session and save manifest
GET	/api/session-info	Get current session info

Keyboard Shortcuts

Key	Action
Space	Start/Stop recording
Enter	Save and go to next phrase

Tech Stack

• Backend: FastAPI, pdfplumber, uvicorn
• Frontend: Vanilla JavaScript, Web Audio API, MediaRecorder API
• Package Manager: uv

Requirements

• Python 3.11+
• Modern browser with microphone support (Chrome, Firefox, Edge)

Tips for Quality Recordings

1. Use a quiet environment
2. Speak clearly and at a natural pace
3. Keep consistent distance from the microphone
4. Use a good quality microphone if available
5. Record in a room with minimal echo

License

MIT