A name of its own
Until this month, grexBot was built to wake up to a stock phrase from its speech library. That's a bit like answering to your part number. Now it has a custom wake word, trained for this robot and running on the robot itself. The first model only knew "Hi grexBot." Then "Hey grexBot" made the guest list, and the current test build listens for both.
The trouble with "hi greks"
Wake words have one natural enemy: the almost-word. For grexBot, that's "hi greks." My first two-phrase model caught every Hi and every Hey in 100 generated test clips of each. It also jumped at "hi greks" 24 times out of 100, so it got rejected. The next round trained hard on that exact sound-alike and cut it to 2 out of 100.
Say hello, then watch some TV
Then it was the robot's turn. In a two-minute live test, I said the two greetings about 20 times in total and didn't notice a single miss. The audio pipeline didn't drop a thing. Next came about ten minutes of TV with nobody saying its name. Zero false wakes. Whatever was on, grexBot was not convinced it was being addressed.
What it does with its name, for now
Hearing its name just gets logged. This test build can't record or upload notes, and wake detection doesn't identify who's speaking. Over in hardware news, the speaker passed its first beep test on a separate bench setup. Replies still play through your phone or computer for now. Next on the list: counting hits and misses at two distances, and a full hour of background noise.