Home Assistant’s experimental Voice Support work tackles wake phrases that go unanswered and commands that lose their first word. Michael Hansen adds a second opinion when the initial verifier rejects a possible wake phrase, and places the start of speech recognition at the detected end of that phrase. Home Assistant also enables the buffered-audio request left off in yesterday’s preview coverage, but only when Voice Support can verify every active wake word.
Hansen’s measurements use 6,188 recorded audio streams from four speakers, including two children, run through Home Assistant’s voice pipeline. Among wake phrases the detector had already caught on the measured channel, one child’s verification rate rises from 28.2% to 88.1%. There is a cost: 21.9% of synthesized similar-sounding phrases that triggered the detector pass verification, compared with 10.1% before.
In the same recorded evaluation, speech recognition loses the command’s first word in 23% of accepted wake phrases, down from 30%. The work also fixes a canceled verification leaving its answer behind for the next request. These are preview results: the app has not been tested on a real device, and its ARM64 images have not been tested.