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 is a separate Docker project under:

gpu-worker/

The main Censorarr host sends extracted 16 kHz audio to the worker. The worker runs Faster-Whisper and returns word timestamps. Detection, review logic, and final remuxing remain on the main Censorarr server.

GPU worker requirements

• Linux host/VM/LXC with Docker
• NVIDIA GPU
• working NVIDIA driver
• NVIDIA Container Toolkit
• network reachability from Censorarr to worker TCP 9000

Install the worker

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

Edit:

ASR_WORKER_TOKEN: "CHANGE_ME_TO_THE_SAME_LONG_RANDOM_TOKEN_AS_MAIN_CENSORARR"

Use a long random token.

Start:

docker compose up -d --build

Check:

docker compose ps
docker compose logs --tail=50

Test the worker locally

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

A healthy worker returns JSON and reports CUDA availability.

Configure the main Censorarr instance

In the UI, configure the remote GPU worker with:

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

Backend choices:

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

Token protocol

Current Censorarr uses:

X-Censorarr-Token

An old pre-rename worker that still expects the legacy header will return:

Invalid worker token

even when the literal token value is correct.

If upgrading an older worker, make sure its worker.py is from the current Censorarr release.

Default GPU settings

The shipped worker defaults are tuned for GTX 10-series / Pascal compatibility:

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

Long WAVs are processed in bounded chunks to avoid large host-RAM spikes.

Network troubleshooting

From the main Censorarr host:

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

If this cannot connect:

• verify worker container is running
• verify port 9000:9000
• verify firewall/routing
• verify correct IP
• verify token

No CUDA device visible

If the worker says no CUDA device is visible, verify NVIDIA Container Toolkit and Docker GPU passthrough before troubleshooting Censorarr itself.