Whisperlater 🎙️

A Dockerized, self-hosted web platform for high-performance transcription and translation. Leveraging faster-whisper (CTranslate2), Whisperlater provides professional-grade subtitles with optimized readability.

🚀 Key Features

• GPU Accelerated: Optimized for NVIDIA GPUs using faster-whisper and CTranslate2.
• Subtitle Safety Cap: Each subtitle cue is limited to a maximum of 15 words.
• Readability Pass: Automatic word-wrapping (42 chars) and minimum display duration (1.2s) for better viewer experience.
• Smart VAD: Built-in Voice Activity Detection (Silero VAD) to eliminate hallucinations in silent audio segments.
• Suppression Defaults: Uses suppress_tokens=[-1,50364] to reduce repetitive hallucinations.
• Long Media Support: Media longer than 4 hours is split into 4-hour chunks for stable processing.
• Modern Web UI: Responsive Vue 3 + Tailwind CSS interface with advanced job configuration.
• Persistence: Mountable volumes for permanent storage of outputs and temporary uploads.

🛠️ Tech Stack

• Backend: FastAPI, Uvicorn, Python 3.10
• Engine: faster-whisper, ctranslate2
• Frontend: Vue.js 3 (CDN), Tailwind CSS (CDN)
• Containerization: Docker with NVIDIA CUDA 12.2

📋 Prerequisites

• Linux OS (Ubuntu 22.04 recommended)
• NVIDIA GPU with 8GB+ VRAM (for large-v3)
• Docker & NVIDIA Container Toolkit installed

📥 Installation & Quick Start

1. Clone the Repository

git clone https://github.com/jeric-n/Whisperlater.git
cd Whisperlater

2. Build the Docker Image

docker build -t whisperlater -f docker/Dockerfile .

3. Run the Container

docker run -d \
  --name whisperlater \
  --gpus all \
  -p 8000:8000 \
  -v $(pwd)/outputs:/app/outputs \
  -v $(pwd)/temp:/app/temp \
  --restart unless-stopped \
  whisperlater

Access the UI at http://localhost:8000.

⚙️ Configuration

Model

Whisperlater uses a single model:

• whisper-large-v3 (internally mapped to large-v3)

Runtime Defaults

• Temperature: 0.0,0.2,0.4,0.6,0.8
• Suppress tokens: [-1,50364]
• No-repeat n-gram guard: 3 for translate tasks (0 for transcribe)
• Long-media split threshold: 4 hours
• Long-media chunk size: 4 hours
• Max subtitle words on screen: 15 per cue

Volumes

• /app/outputs: Finished .srt and .txt files are stored here.
• /app/temp: Temporary upload storage (auto-cleaned after processing).

🧪 Development & Testing

Run the app locally:

1. Install dependencies: pip install -r requirements.txt
2. Start server: uvicorn app.main:app --host 0.0.0.0 --port 8000

📄 License

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