Not that I disagree with your conclusion because there’s an even simpler way to check if an app is listening: iOS and Android will tell you the mic is being used… Anyway, we do have always-on NNs listening for keywords (“Siri,”, “Hey google”, “Alexa”) so I agree that full ass voice transcription like whisper will run like dogshit on your phone they can certainly run a much much lighter model to pick up a handful of keywords.
I most definitely trust Apple, Google, and Amazon to not use the backdoor listening devices the NSA likely had them put in all phones.
You are probably right though. I doubt this anything like they are marketing it. It’s likely location tracking, flip on mic near businesses with services but don’t record till input reaches -12db, transcribe, parse keywords against list and report back. This whole process is incredibly low power on a phone and the only packet you would probably see leave the device is small and probably pretty inocuous looking.
To Camdat’s point, a general transcription is definitely not low power even if you have some kind of gating on when it transcribes. Obviously Apple and Google and Samsung and whoever makes the phone can turn on the mic without you knowing, otherwise how would their voice assistant work, but Apple probably isn’t letting Facebook have access to the mic without throwing something up on the status bar.
I just did a rough whatsapp text to speech message using a tasker script log and monitor battery power usage. I had it rounding to 4 decimal places of a percentage and saw 0 percent change. Take my rather unscientific test as you will. I think the trick is to not do it continuously and leverage location and volume detection events. I think a few years ago this may have been a lot more heavy on the battery but anything made within the last 2-3 years will not have much of a problem.
Sure this is definitely true. I should clarify that single-word NNs do run on-device all the time, but those require specialized models that are trained only on those keywords. Once those models trigger they need to send everything else to the cloud.
I agree. If I was going to do something like this for advertising though I wouldn’t really care too much about what people were saying so instead I’d just listen for some limited set of keywords (maybe for some of my top paying advertisers) and serve ads for keywords that hit recently. Keep it all on device until an ad actually needs to be served.
Not that I disagree with your conclusion because there’s an even simpler way to check if an app is listening: iOS and Android will tell you the mic is being used… Anyway, we do have always-on NNs listening for keywords (“Siri,”, “Hey google”, “Alexa”) so I agree that full ass voice transcription like whisper will run like dogshit on your phone they can certainly run a much much lighter model to pick up a handful of keywords.
I most definitely trust Apple, Google, and Amazon to not use the backdoor listening devices the NSA likely had them put in all phones.
You are probably right though. I doubt this anything like they are marketing it. It’s likely location tracking, flip on mic near businesses with services but don’t record till input reaches -12db, transcribe, parse keywords against list and report back. This whole process is incredibly low power on a phone and the only packet you would probably see leave the device is small and probably pretty inocuous looking.
To Camdat’s point, a general transcription is definitely not low power even if you have some kind of gating on when it transcribes. Obviously Apple and Google and Samsung and whoever makes the phone can turn on the mic without you knowing, otherwise how would their voice assistant work, but Apple probably isn’t letting Facebook have access to the mic without throwing something up on the status bar.
I just did a rough whatsapp text to speech message using a tasker script log and monitor battery power usage. I had it rounding to 4 decimal places of a percentage and saw 0 percent change. Take my rather unscientific test as you will. I think the trick is to not do it continuously and leverage location and volume detection events. I think a few years ago this may have been a lot more heavy on the battery but anything made within the last 2-3 years will not have much of a problem.
Whatsapp is sending your audio to the cloud to handle transcription. This is not an accurate test because it is not an on-device process.
No, my keyboard is doing it locally. Futo.
Sure this is definitely true. I should clarify that single-word NNs do run on-device all the time, but those require specialized models that are trained only on those keywords. Once those models trigger they need to send everything else to the cloud.
I agree. If I was going to do something like this for advertising though I wouldn’t really care too much about what people were saying so instead I’d just listen for some limited set of keywords (maybe for some of my top paying advertisers) and serve ads for keywords that hit recently. Keep it all on device until an ad actually needs to be served.