Voice lab

1. Record samples

Your wake phrase

Say the phrase 10 times per round, each in a slightly different style (normal, quiet, quick, with a command, from across the room). Do a few rounds in different rooms and with each microphone. Aim for 20 or more.

Someone else (optional)

Another person says the phrase 5 times, to check that “Only my voice” ignores them.

Background audio

Play TV, music or a video call, or let people talk, for 2, 5 or 10 minutes. Used to measure how often Dexibi would wake by mistake. Aim for 10 minutes or more.

While you record, the assistant stops listening so it doesn’t react to the samples.

2. Evaluate

Evaluate runs the on-device wake-word model and the voice check over every sample, at every setting:

• Wake word: the share of your phrases detected at Low, Medium and High sensitivity, and false wakes per hour of background audio.
• Only my voice: how many of your samples are accepted, and how many other voices and background sounds get through, at Relaxed, Balanced and Strict.
• Recommendation: the most sensitive wake setting with at most one false wake per hour, whether On-device only is reliable enough for you (95% or better), and the strictest voice setting that still accepts you 95% of the time. Apply sets them.

3. Your commands

Paste the commands you use most, one per line. Dexibi shows which run instantly on your PC and which are planned by the AI agent. Nothing is executed.

Privacy

Recordings are saved only on your PC, in a folder you can open from the Voice lab. They’re never uploaded. Delete all removes them, and uninstalling offers to remove them too.