A local voice server with its speech model already downloaded should not need the internet to start. OHF Voice’s Wyoming Faster Whisper server tries its cached models first, avoiding a Hugging Face connection that could stall startup and send an isolated container into a restart loop.
Michael Hansen applies the approach across the supported recognition backends. A missing model can still trigger a download; the explicit local-files-only option instead fails immediately when the model is absent. The correction also makes FunASR respect the chosen model directory.