Cloud-Based OBS Captions — No CPU, No GPU

Why local OBS caption plugins hurt your stream

The most popular free OBS caption plugin (LocalVocal) runs OpenAI Whisper locally on your GPU. That means every word you speak triggers Whisper inference on your machine. Typical impact:

• 5-20% GPU utilization during active streaming
• 500MB-2GB VRAM reserved for the model
• 5-10% CPU for audio preprocessing and result rendering
• Extra fan noise from sustained GPU load

For streamers already pushing the GPU with games + OBS encoding, this matters. You either drop encoding bitrate, reduce streaming resolution, or get worse Whisper transcription accuracy because the model has less budget. It is a triangle: game quality, stream quality, caption quality — pick two.

Why cloud-based is different

StreamTranslate moves the speech recognition off your machine entirely. The flow:

1. Your browser (running as an OBS browser source) captures mic audio via MediaRecorder
2. Audio streams over WebSocket to our servers
3. Deepgram Nova-3 runs the transcription on dedicated GPU infrastructure
4. Translation runs in parallel (DeepL + Google + Azure race for fastest result)
5. Result streams back to the browser overlay

Your machine's job: encode audio, render the overlay text. Both take near-zero CPU. The heavy work happens in the cloud where dedicated GPU servers can crunch through hundreds of streams in parallel without affecting any one streamer's machine.

Resource comparison: local vs cloud

When to pick which

Pick LocalVocal if: you have a beefy GPU (RTX 4070+), do not need translation accuracy beyond Whisper's built-in, and want totally free / offline operation.

Pick StreamTranslate if: your GPU is already busy with games, you want sub-500ms latency, you need 30+ languages with multi-provider translation, or you cannot afford to have your plugin break on the next OBS update.

Frequently asked questions