Corey JamesRecognition on my own hardware
Helix holds spoken conversations in a room or on a Microsoft Teams call, and I wanted the speech it hears recognised on my own hardware. the speech recognition model is the recognition model; the engineering is how the voice engine uses it, and above all what it does with an uncertain result. the speech recognition model is an open model served through the open-source its C++ runtime; the integration, the confidence policy and the deployment are mine.
The engine's transcriber is an interface with two implementations: the speech recognition model in process through its C++ runtime, and a client for a the speech recognition model server on a GPU host that takes WAV audio and returns segments. The two deployments swap without touching conversation logic. In the Teams configuration the small English model runs on a Blackwell GPU next to neural speech synthesis, so a call's audio stays on my hardware. Decoding is deterministic, so the same audio always gives the same text, which makes a conversation reproducible when I look back at it.
Silent rather than wrong
Speech models invent fluent text out of noise and silence. So every recognised segment carries a confidence measure, and segments under the threshold are dropped. If nothing in a turn survives, the engine produces an empty turn, because an empty turn beats invented text. Configuration follows the same rule: a missing recogniser is a loud failure with no silent fallback.
A digital employee that answers words nobody said is worse than one that stays quiet.
The call that first proved the chain, ringing a Teams number and talking to a digital employee whose recognition, reasoning and voice all ran on local GPUs, is A self-hosted colleague in a Teams call. The same preference for an honest gap over a fabricated answer runs through A meeting attendee that will not fake it.