Gregory's Server@Unixbigot Good to know thank you!
|
13 comments
 @diazona @shannonpersists I recently considered getting a custom wakeword. The hardware vendor says they will need at least 20000 recordings from at least 500 speakers of all ages under various (specified) conditions. It would take several weeks to train the model and cost <contact sales> dollars. Reader, I did not contact sales because by that point I suspected the fifty four cents I had to hand would not suffice.  @Unixbigot @diazona @shannonpersists Oh, it depends on how skilled a salesdroid you get. You may find yourself happily dancing away waving a signed contract that accepts your 54c... as a first installment.  @Unixbigot @shannonpersists 😂 I suspect you are right. I used to work at a voice recognition company, and one of the services we offered was custom wakewords. The kinds of clients that took us up on that offer had (many-)multi-million dollar contracts. Of course they were paying for a lot more than the custom wakeword, but the cost of training a wakeword model is definitely substantial.  @Unixbigot @diazona @shannonpersists @diazona @shannonpersists @Unixbigot @diazona @shannonpersists @Unixbigot @bornach @diazona @shannonpersists @Unixbigot @shannonpersists @bornach @diazona @Unixbigot The future is already here. Just unevenly distributed. @diazona @freemin7 @bornach @Unixbigot It seems like all it would take is one spy and a rogue middle-aged mechanic!💥🌎 |
@shannonpersists @Unixbigot And on top of that, the wake word detection system doesn't know your specific voice (at least not until quite recently), so it has to listen for the wake word in many different accents, with or without background noise, whispered close up or shouted from across the room, and so on. People have managed to make these systems quite good, but they do need to cast a fairly broad net, so to speak. Which means there's always going to be some chance that they misinterpret other noises as the wake word.
@shannonpersists @Unixbigot And on top of that, the wake word detection system doesn't know your specific voice (at least not until quite recently), so it has to listen for the wake word in many different accents, with or without background noise, whispered close up or shouted from across the room, and so on. People have managed to make these systems quite good, but they do need to cast a fairly broad net, so to speak. Which means there's always going to be some chance that they misinterpret other...