transcribe-Web

Whisper-based transcription with both CLI and Web App modes.

• Browser upload flow for audio/video files
• Docker-first startup with CUDA acceleration
• Optional Cloudflare Tunnel sidecar for public access

Languages

• English (this file)
• 繁體中文: README.zh-TW.md

---

Web App Overview

The web app is implemented with FastAPI and serves both frontend and API from one process.

Architecture:

Browser <--HTTPS--> Cloudflare Tunnel <--> Docker (FastAPI:80) <--> transcribe() API
                                                       |
                                                    CUDA GPU

API Endpoints

• GET / — Embedded frontend page
• POST /api/transcribe — Upload file + create transcription job
• GET /api/jobs — List jobs
• GET /api/jobs/{job_id} — Poll job status
• GET /api/jobs/{job_id}/download/{filename} — Download outputs

Output files

A completed job may include:

• out.srt
• out.vtt
• out.txt
• out.json
• speaker.json (only when supported by backend/options)

---

Quick Start

Local Python (Web mode)

./install
source activate
transcribe-anything-web

Open: http://localhost:80

Docker Compose (Recommended)

1. Create .env in project root:

CLOUDFLARE_TUNNEL_TOKEN=your-tunnel-token-here

2. Build and run:

docker compose up --build -d

3. Open local web app:

• http://localhost:8092

4. Stop:

docker compose down

---

Docker Build / Startup / Deployment Notes

What the image does at build time

The Docker image prewarms backend environments to reduce first-request latency:

• transcribe-anything-init-cuda
• transcribe-anything-init-insane

The image defaults to web mode:

• ENTRYPOINT ["/app/entrypoint.sh"]
• CMD ["--web"]

What entrypoint.sh does at runtime

• Configures CUDA runtime library paths
• Runs shared-library checks
• Starts web server for --web

Current Compose services

docker-compose.yml defines:

• transcribe
  • Built from local Dockerfile
  • Port mapping: 8092:80
  • Volume: transcribe-data:/app/data
  • Env:
    • MAX_UPLOAD_SIZE_MB=0
    • NVIDIA_VISIBLE_DEVICES=all
• cloudflared
  • Image: cloudflare/cloudflared:latest
  • Command: tunnel run
  • Env: TUNNEL_TOKEN=${CLOUDFLARE_TUNNEL_TOKEN}

Cloudflare Tunnel setup

1. Create a Tunnel in Cloudflare Zero Trust.
2. Put token into .env as CLOUDFLARE_TUNNEL_TOKEN.
3. Configure public hostname service target to:
   • http://transcribe:80
4. Restart stack:

docker compose down
docker compose up -d

Upload limits

• App-side upload limit is controlled by MAX_UPLOAD_SIZE_MB (default unlimited; set 0, none, or unlimited for no limit).
• Cloudflare edge upload limits may still apply on public URLs.
• For very large uploads, use local endpoint http://localhost:8092.

---

CLI / Python API (still supported)

CLI examples:

transcribe-anything video.mp4 --device cpu
transcribe-anything video.mp4 --device cuda
transcribe-anything video.mp4 --device insane --batch-size 8
transcribe-anything video.mp4 --device mlx

Python API:

from transcribe_anything import transcribe

transcribe(
    url_or_file="video.mp4",
    output_dir="output_dir",
    task="transcribe",
    model="small",
    device="cuda",
)

---

Development commands

• Setup: ./install && source activate
• Tests: ./test
• Lint: ./lint --no-ruff
• Clean: ./clean

Acknowledgements

• Originally derived from transcribe-anything

License

BSD-3-Clause