Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

My echo absolutely became less accurate month after month. Simple commands like “turn off the lights” would be 50/50. Very low-value technology and it certainly gives one pause about the feasibility of AI when something so basic with so much capital and years of effort is still fairly poor.


I’ve been wondering if I’ve been going crazy. It’s gotten way worse for me too. I have to ask 3-4 times to turn the lights off sometimes.


Research is going strong. The Whisper model, recently open sourced, is great for ASR and runs on the edge. https://openai.com/blog/whisper/


Is there a recommended dev platform to test this model? e.g. Google/Coral TPU, Nvidia Jetson Nano, Rockchip RK3568 NPU, iPhone/Android NPU, ..?


Coral and Jetson are not very powerful actually, you can't run even medium model in realtime. https://github.com/openai/whisper/discussions/417

And there is latency issue too, you won't get response very fast. If you look for <0.5s latency or the answer, you need to try something like Vosk probably.


I’ve been assuming physical factors like dust was degrading the microphone quality


Also getting worse for me. I upgraded from my dot to a regular gen 5 echo, but no, it's still bad.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: