Raising local LLM voice command accuracy: aliases, exposure and instructions

Say the same sentence twenty times and note the outcome. Whether it was misheard, turned on another device, or picked the right one and did nothing each points to a different fix.

Mishearing is the most common. Check microphone placement and sentence length first, then aliases and instructions.

Mixing 'living room light' and 'main lamp' makes the model interpret each differently. Adding an alias on the Expose tab under Settings > Voice assistants ties both names to one device.

Aliases also work for areas and floors. In a house that mixes Korean and English, both names make the same device reachable in either language.

An alias does nothing if the entity is not exposed. The conversation agent can only see entities exposed through the Assist API¹, so check that frequently used lights, switches and climate devices are on the list, and leave locks and garage doors off.

A very long list makes the choice harder for the model. Keep the devices each room uses and drop seasonal ones for a while.

Put naming rules and the answer format in the instructions: say only the device and its state when switching, and ask rather than act when unsure.

Long instructions delay the first response; short ones invite wrong actions. After adding each rule, rerun the twenty-sentence test and see whether failures dropped.

If failures still exceed half after names and lists are tidy, a larger model is worth trying. If most failures are wrong-device picks, the list is the problem, not the model.

A large model answers slowly and the conversation stalls. Keep device control on a small model and send long questions through text input instead.