Transcription and GPU Worker

Transcription & GPU Worker

Censorarr supports local CPU transcription or an optional remote NVIDIA GPU Worker.

Local CPU

Default-style configuration:

whisper:
  backend: local
  model: small
  device: cpu
  compute_type: int8
  language: en

This is the simplest backend for a first-run test.

Remote GPU Worker

The worker receives extracted audio from the main Censorarr instance, runs Faster-Whisper on an NVIDIA GPU, and returns word timestamps. Detection, review logic, and final media remuxing remain on the main Censorarr server.

You can run the worker three ways:

• Native Windows 11 x64 installer
• Native Debian/Ubuntu x86_64 .deb
• Docker using the existing gpu-worker/ compose project

All three use the same API and X-Censorarr-Token protocol.

Native Windows GPU Worker

Download from GitHub Releases:

Censorarr-GPU-Worker-Setup-X.Y.Z.exe

The installer automatically:

1. installs the worker
2. generates a strong worker token
3. downloads the pinned NVIDIA CUDA 12 / cuBLAS / cuDNN runtime files
4. verifies their SHA-256 hashes
5. stores the NVIDIA runtime privately under Censorarr GPU Worker data
6. creates an automatic startup task
7. starts the API on TCP 9000

The NVIDIA display/compute driver is the one prerequisite that is not installed automatically.

Configuration is stored at:

C:\ProgramData\CensorarrGPUWorker\worker.env

The installer opens this file at the end so you can copy ASR_WORKER_TOKEN.

If the NVIDIA driver is installed later, run:

& "C:\Program Files\Censorarr GPU Worker\CensorarrGPUWorker.exe" --install-runtime
& "C:\Program Files\Censorarr GPU Worker\CensorarrGPUWorker.exe" --start-service

Native Debian / Ubuntu GPU Worker

Download:

Censorarr-GPU-Worker-X.Y.Z-linux-amd64.deb

Install:

sudo apt install ./Censorarr-GPU-Worker-X.Y.Z-linux-amd64.deb

Locations:

Item	Location
Application	/opt/censorarr-gpu-worker
Configuration/token	/etc/censorarr-gpu-worker.env
Models	/var/lib/censorarr-gpu-worker/models
Private NVIDIA runtime	/var/lib/censorarr-gpu-worker/runtime
Service	censorarr-gpu-worker.service
API	http://HOST:9000

Show the token:

sudo /opt/censorarr-gpu-worker/CensorarrGPUWorker --show-token

Service commands:

sudo systemctl restart censorarr-gpu-worker
systemctl status censorarr-gpu-worker
journalctl -u censorarr-gpu-worker -f

If the driver is installed after the package:

sudo /opt/censorarr-gpu-worker/CensorarrGPUWorker --install-runtime
sudo systemctl restart censorarr-gpu-worker

What the native installer downloads

The native installer does not install the full CUDA development toolkit and does not modify an existing global CUDA setup.

At build time Censorarr creates a platform-specific manifest containing exact NVIDIA package versions, file URLs, sizes, and SHA-256 hashes. During installation it downloads those exact NVIDIA-maintained PyPI wheels and verifies the hashes before extracting the runtime libraries locally.

The pinned root runtime components are:

• CUDA 12 runtime
• cuBLAS for CUDA 12
• cuDNN 9 for CUDA 12

Any NVIDIA package dependencies are resolved into the manifest at build time as well.

A working NVIDIA driver is still required.

Docker GPU Worker

Docker remains supported:

git clone https://github.com/leestow/Censorarr.git
cd Censorarr/gpu-worker

Set a long random token in docker-compose.yml:

ASR_WORKER_TOKEN: "CHANGE_ME_TO_THE_SAME_LONG_RANDOM_TOKEN_AS_MAIN_CENSORARR"

Then:

docker compose up -d --build

Test the worker

From the worker host:

curl -H 'X-Censorarr-Token: YOUR_TOKEN' http://127.0.0.1:9000/health

From the main Censorarr host:

curl -H 'X-Censorarr-Token: YOUR_TOKEN' http://GPU-WORKER-IP:9000/health

A healthy worker should report one or more CUDA devices.

Configure the main Censorarr instance

In Settings → Setup Wizard or Transcription settings:

URL:   http://GPU-WORKER-IP:9000
Token: exactly the same ASR_WORKER_TOKEN used by the worker

Backend choices:

• remote — GPU worker only
• auto — GPU worker with local CPU fallback

Default GPU settings

Model:          small.en
Compute type:   int8_float32
Chunk size:     600 seconds
Chunk overlap:  2 seconds

These defaults remain suitable for GTX 10-series / Pascal cards. Long WAV files are processed in bounded chunks to avoid large host-RAM spikes.

Troubleshooting

Invalid worker token

The main Censorarr token and worker ASR_WORKER_TOKEN must match exactly. Current Censorarr uses:

X-Censorarr-Token

No CUDA device visible

Check:

nvidia-smi

For native installs, also verify the private runtime was installed:

Windows:

& "C:\Program Files\Censorarr GPU Worker\CensorarrGPUWorker.exe" --runtime-plan

Linux:

sudo /opt/censorarr-gpu-worker/CensorarrGPUWorker --runtime-plan

Then rerun --install-runtime if needed.

Cannot connect to port 9000

Verify:

• worker process/service is running
• TCP 9000 is allowed through the host firewall
• the IP/hostname is correct
• Censorarr can route to the worker host