Skip to main content

Hungarian offline voice stack

This fork adds two local providers to IntraCord's Model Configurations screen.

  • Nemotron Local: NVIDIA Nemotron Streaming 3.5 ASR for real-time Hungarian transcription. The server field expects a local, Riva-compatible gRPC address; the Docker default is host.docker.internal:50051.
  • Piper: free local text-to-speech. Available Hungarian voices are hu_HU-anna-medium, hu_HU-berta-medium, and hu_HU-imre-medium. Piper downloads the selected voice to the persistent /data/piper volume on first use and can run offline afterward.

Start the stack​

.\scripts\start_docker.ps1
# Do not start it at the prompt yet. Then run:
docker compose -f docker-compose.yaml -f docker-compose.local-ai.yaml --profile tunnel up --build -d

The application is available at http://localhost:3010.

Local Nemotron server​

The Nemotron adapter uses the Riva/NIM gRPC protocol. Start a local Riva-compatible server containing the nemotron-3.5-asr-streaming-0.6b model on port 50051; the model and its weights then remain entirely local. With Docker Desktop, keep the default host.docker.internal:50051 address. For another address, update the configuration's server field. Set use_ssl=false for a local server without TLS.

The OpenAI LLM provider remains available. Enter your own API key in the OpenAI configuration.