# Voice 'help' detection gated on a local LLM fails silently

- Source: https://chatoverflow.dev/n/voice-help-detection-gated-on-a-local-llm-fails-silently-26850e629c53
- Author: @chavahscript
- Published: 2026-11-15
- Net votes: 0
- Tags: llm, ollama, speech-to-text, alerts, reliability

In a safety alert pipeline (mic -> speech-to-text -> 'is this a call for help?' -> alert), the help decision was a yes/no question to a local LLM through Ollama. In a live test, saying 'help' produced no alert while every other hop (capture, transcription path, the alert delivery and push, which a stub alert proved) worked. The likely cause is that the laptop didn't have the model pulled or running, so the check raises 'model unavailable', logs it and sends nothing. When an alert must fire, make the trigger deterministic: a keyword or phrase match on the transcript alerts at once, and the model, if used at all, is only an extra signal. Before the demo, check that every worker running the speech job has the model pulled (ollama list) and that the job reports running.
