Parakeet TDT - Audio Transcriber

A professional-grade full-stack application for recording audio in the browser and transcribing it locally using NVIDIA's Parakeet ASR model with advanced speaker diarization.

✨ Features

• 🎙️ Live Audio Recording: High-quality audio capture directly from your browser
• 🤖 Advanced ASR: Dual-engine support (NVIDIA Parakeet + OpenAI Whisper)
• 🔊 Speaker Diarization: Automatic identification of speakers using pyannote.audio (~90% accuracy)
• 📨 Live Transcription: Real-time message-based transcript display with smart pause detection
• 💾 Transcript History: Browse, search, and manage all saved transcripts
• 📥 Export Options: Download transcripts as text files or view as JSON
• 📋 Clipboard Support: Copy transcripts with one click
• 🌐 Multi-user: Support for 3+ concurrent transcription users
• 🚀 100% Local Processing: All transcription happens locally, no cloud APIs
• 🎯 Intelligent VAD: Voice Activity Detection with natural pause recognition

Project Structure

.
├── frontend/          # Next.js frontend application
│   ├── app/
│   │   ├── components/
│   │   │   ├── AudioRecorder.tsx
│   │   │   └── TranscriptHistory.tsx
│   │   ├── layout.tsx
│   │   ├── page.tsx
│   │   └── globals.css
│   ├── package.json
│   ├── tsconfig.json
│   └── tailwind.config.js
└── backend/           # Python FastAPI backend
    ├── main.py
    ├── database.py
    ├── schemas.py
    ├── requirements.txt
    └── .env.example

Prerequisites

• Node.js 18+ (for frontend)
• Python 3.9+ (for backend)
• GPU (recommended for faster transcription, NVIDIA CUDA preferred)

Backend Setup

1. Create and activate virtual environment

cd backend
python -m venv venv

# On macOS/Linux
source venv/bin/activate

# On Windows
venv\Scripts\activate

2. Install dependencies

pip install -r requirements.txt

Note: The first time you run the application, Parakeet will download the pre-trained model (~1.5GB), which may take a few minutes.

3. Create environment file

cp .env.example .env

4. Run the server

uvicorn main:app --reload --host 0.0.0.0 --port 8000

The backend will be available at http://localhost:8000

Frontend Setup

1. Install dependencies

cd frontend
npm install

2. Run the development server

npm run dev

The frontend will be available at http://localhost:3000

Usage

1. Open http://localhost:3000 in your browser
2. Click "Start Recording" to begin recording audio
3. Speak into your microphone
4. Click "Stop Recording" when done
5. Wait for the transcription to complete
6. View your transcript, copy it, or download it as a text file
7. Your transcripts are saved in the "Transcript History" section

API Endpoints

POST /transcribe

Upload an audio file for transcription

• Parameters:
  • file: Audio file (WAV, MP3, etc.)
  • title: Optional title for the transcript

GET /transcripts

Get all saved transcripts

GET /transcripts/{transcript_id}

Get a specific transcript by ID

DELETE /transcripts/{transcript_id}

Delete a transcript

WS /ws/transcribe

WebSocket endpoint for streaming transcription (future enhancement)

Troubleshooting

"Unable to access microphone"

• Make sure your browser has microphone permissions
• Check your browser's privacy settings
• Try using HTTPS (required for microphone access on non-localhost)

"Model not loaded" error

• Make sure you have internet connection for the first run (to download the model)
• Check that your GPU has enough memory (recommended: 4GB+)
• For CPU-only, transcription will be slower

Port already in use

• Frontend: Change port in npm run dev -- -p 3001
• Backend: Change port in uvicorn main:app --port 8001
• Update CORS origins in backend/main.py accordingly

Performance Notes

• GPU: With NVIDIA GPU, transcription typically takes 2-5 seconds per minute of audio
• CPU: Without GPU, transcription may take 30+ seconds per minute of audio
• The Parakeet model requires about 1.5GB of disk space and ~2GB of RAM

Quick Start

Fast Setup (macOS/Linux)

./start.sh

Manual Setup

Backend:

cd backend
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
uvicorn main:app --reload

Frontend:

cd frontend
npm install
npm run dev

Then open http://localhost:3000 in your browser.

Documentation

• CLAUDE.md - Detailed architecture, command reference, and implementation notes
• CONTRIBUTING.md - How to contribute to this project
• GETTING_STARTED.md - Detailed setup instructions
• API.md - API endpoint documentation

Architecture

Speaker Diarization

The system uses pyannote.audio to identify and track speakers with ~90% accuracy:

• Analyzes complete audio after recording
• Automatically detects number of speakers
• Labels segments with speaker attribution
• Graceful fallback if pyannote unavailable

Voice Activity Detection (VAD)

Smart pause detection breaks natural speech into manageable segments:

• 0.3-0.5s pause triggers transcription
• 1.5+ seconds confirms end of phrase
• 15+ second buffer forces send to prevent huge blocks

Multi-user Support

• Up to 3 concurrent transcriptions
• Configurable worker threads
• WebSocket support for real-time updates
• PostgreSQL or SQLite database backend

Performance

Scenario	Time	Hardware
1 min audio (GPU)	5-10s	NVIDIA GPU + 8GB RAM
1 min audio (CPU)	5-10 min	CPU-only
Multi-user (3 concurrent)	Sequential	~3x slower per user

Environment Setup

Optional: HuggingFace Token

For better speaker diarization model access:

# Create token at https://huggingface.co/settings/tokens
# Add to .env:
HUGGINGFACE_TOKEN=hf_your_token_here

Database Configuration

# Default: SQLite (auto-created)
# For PostgreSQL:
DATABASE_URL=postgresql://user:pass@localhost/transcriber_db

Future Enhancements

• Support for multiple languages
• Confidence scores for transcribed text
• Batch transcription
• Audio file upload without recording
• Real-time speaker identification display
• Custom speaker name assignment
• Transcript search and filtering
• Timestamp-based playback
• Speaker demographics (experimental)

License

This project uses open-source components:

• Frontend: Next.js (MIT)
• Backend: FastAPI (MIT)
• ASR Model: Nvidia Parakeet (Apache 2.0)

Support

For issues or questions:

1. Check the troubleshooting section
2. Verify all prerequisites are installed
3. Check that both backend and frontend servers are running
4. View browser console and backend logs for error messages