4 ms·

I'm pretty sure that speech-to-text is still a major issue. Most solutions available right now are all too heavy to run on-device, there's a reason only wakewor

I'm pretty sure that speech-to-text is still a major issue. Most solutions available right now are all too heavy to run on-device, there's a reason only wakeword detection is done on device and the audio is actually transcribed on the cloud. Even the one the parent linked, OpenAI Whisper, seems to require a beefy machine to run. When we get models/apis that can run on a raspberry pi in real time then we can say the issue is ecosystem access. Right now i'm unable to build my own custom assistant, simply because I have not yet found a speech-to-text model that can run on my low power devices - even the one on my top-of-the-line android phone needs to call out to Google for recognition!

shagie 4y ago
If you've got an iPhone, kick it into airplane made and open up "notes" and tap the microphone to transcribe. It even gets "one third plus one half is five sixths" transcribed as `1/3 + 1/2 is 5/6` and "four feet two inches" becomes `4'2"` without a network connection. So I believe its doable... I also believe that Apple spent bit to make it that way. Note that the voice to text on the device cuts down on data use (I believe - I haven't sniffed the traffic to verify) and cloud side processing (its working on text rather than audio - its a few pennies less per request). But beyond that... every home voice assistant that I've seen (other than Apple, Amazon, and Google) has worked on pre-programed phrases that need to be spoken rather than the more free form way people speak... and again, that's even without trying to translate that into other infrastructure beyond the device. I'd honestly be happy/impressed to have a RPi based system that can handle taking "create a reminder for next Wednesday at 9:30 am to thaw the turkey" and have that feed into other useful systems. I'd be very impressed if I could do a "create a reminder" (what is the reminder for) "go to parents" (when should I remind you) "Day after thanksgiving" (what time on Friday should I remind you) "half past one" (is that one thirty in the morning or in the afternoon) "afternoon" -- this works with Alexa. Note the recognition of holidays, the 'wizard' style entry and the alternate time format.

shagie 4y ago

If you've got an iPhone, kick it into airplane made and open up "notes" and tap the microphone to transcribe. It even gets "one third plus one half is five sixths" transcribed as `1/3 + 1/2 is 5/6` and "four feet two inches" becomes `4'2"` without a network connection. So I believe its doable... I also believe that Apple spent bit to make it that way. Note that the voice to text on the device cuts down on data use (I believe - I haven't sniffed the traffic to verify) and cloud side processing (its working on text rather than audio - its a few pennies less per request). But beyond that... every home voice assistant that I've seen (other than Apple, Amazon, and Google) has worked on pre-programed phrases that need to be spoken rather than the more free form way people speak... and again, that's even without trying to translate that into other infrastructure beyond the device. I'd honestly be happy/impressed to have a RPi based system that can handle taking "create a reminder for next Wednesday at 9:30 am to thaw the turkey" and have that feed into other useful systems. I'd be very impressed if I could do a "create a reminder" (what is the reminder for) "go to parents" (when should I remind you) "Day after thanksgiving" (what time on Friday should I remind you) "half past one" (is that one thirty in the morning or in the afternoon) "afternoon" -- this works with Alexa. Note the recognition of holidays, the 'wizard' style entry and the alternate time format.