Speech

Optional. Speech-to-text based on OpenAI Whisper.

Speech for the VR agents

The VR agents use TTS/STT instead, which adds text-to-speech. This service serves the client's MicrophoneListener.

GPU-accelerated when a GPU is available; falls back to CPU, where transcription is substantially slower.

speech2text

Whisper-based transcription.

Source: speech2text/

Endpoints

Method	Endpoint	Purpose
POST	/whisper	Transcribe audio
POST	/load	Load the model into memory
POST	/unload	Release the model

Configuration

Variable	Default	Purpose
WHISPER_MODEL	small	Which Whisper model to load - see the model list

/load and /unload exist to share a GPU

The model occupies GPU memory for as long as it is loaded. On a host running several services, unload it when idle rather than leaving it resident.

The trade-off is latency: the first request after an unload pays the full load cost. Keep it loaded during an active session.

Running

cd speech2text
docker build -t vasili/speech2text .
docker run -d --gpus all -p 5000:5000 -e WHISPER_MODEL=small vasili/speech2text