2 ms·
Every Google Home device is already running an ML model to do speech recognition to recognize the "hey Google" wake word, so sooner than 10 years. The Raspberry
by fragmede 3y ago
Every Google Home device is already running an ML model to do speech recognition to recognize the "hey Google" wake word, so sooner than 10 years. The Raspberry Pi Zero is a particularly underpowered device for this. Doing it on the Coral TPU accelerator plugged into a pi zero would take less than 30 mins. Doing it on an iPhone 15 would take less time. Doing it on a Pixel 8 would be faster. Not to diminish getting it to work on a Pi Zero, but that future is already here, just as soon as we figure out what to do with them.
https://coral.ai/products/accelerator/ https://coral.ai/products/accelerator/
- dontwearitout 3y agoThere's an ocean of difference between optimizing for a single wakeword and the class of models that are taking off today. I'm excited for more on-board processing, because it will mean less dependency on the cloud.
- hereonout2 3y agoSiri came out 12 years ago, wake words are probably a bad example. Better examples are the magic eraser on my 2 generation old pixel phone or the fact llama2 runs genuinely fast on a Mac mini
- joegibbs 3y agoSiri’s wake word stuff is also terrible, she gets constantly activated whenever I have my Apple Watch near running water, frying food or anything else that makes a white noise-type sound.
- fragmede 3y agoI'm not going to argue that there isn't a difference when going from 0 -> 1 and 1 -> 10, or in this case, from 1.5b (Whisper-large) -> 1.7t parameters (gpt-4). But it's not like we don't know how to do it, so it won't take 10 years to get there.