https://machinelearning.apple.com/research/hey-siri
> This process not only reduces the probability that "Hey Siri" spoken by another person will trigger the iPhone, but also reduces the rate at which other, similar-sounding phrases trigger Siri.
Apple:
- Only ever uploads anything to the cloud after a *successful* wake word is detected on-device (and, critically, lets you know with a voice confirmation).
- "On most Echo devices" checks on-device against known commercials (the point of this article).
- "In the cloud: Every audio request to Alexa that starts with a wake word is checked..." — successful or not. Meaning, all audio requests are streamed, then problematic wake words are filtered, then the phrase is acted on.
> Ideally, a device will identify media audio using locally stored fingerprints, so it does not wake up at all. If it does wake up, and we match the media event in the cloud, the device will quickly and quietly turn back off.
i.e. if a media match is not detected locally, the audio (all audio) is next sent to the cloud for screening & possible action.
Everything here on the Amazon side is about detected false-positives, on-device or not, and nothing is about protecting the user's privacy.
> is very possible that anything could be matched (i.e. sent) in the cloud.
You misread it. Only after a wakeword (Alexa) is anything sent, and then it gets double checked. Apple doesn't see to have any double check at all - that doesn't make it better.
I'd bet that it totally omits sending any audio for exclusions processed on device, and that the article is not ideally worded. It would be true that some devices may not be able to perform on device exclusion processing (perhaps some of the oldest ones). Further cloud processing would use any information not available to local processing (like sound signatures from commercials no longer running, or the unknown media processing described).
> We compare the distances to the reference patterns created during enrollment with another threshold to decide whether the sound that triggered the detector is likely to be "Hey Siri" spoken by the enrolled user.
> This process not only reduces the probability that "Hey Siri" spoken by another person will trigger the iPhone, but also reduces the rate at which other, similar-sounding phrases trigger Siri.
Contrast with Amazon, where the mere suspected presence of the wake word is enough to get it sent to the cloud for further analysis.
My overall point is that Amazon is much more indiscriminate about assuming you are talking to it and sending stuff to the cloud to act on (which sometimes includes discarding). Whereas Apple will only send what it believes to be a command.
Amazon will continue to build up a database of sent audio in order to improve this cloud-based double-check of the wake word, whereas Apple will discard a command that it didn't know how to act on, never needing to improve a cloud-based wake word refinement.
The stuff you are writing is just not true.
> This process not only reduces the probability that "Hey Siri" spoken by another person ... Contrast with Amazon
Siri is personal on a single phone, while Alexa is meant to reply to anyone in the room.
You need to figure out why you have this blind spot for Apple, it's causing you to misunderstand things.
But I still think Apple's approach is radically more aggressive on the "what it believes to be a wake word" angle. And the fact that wake words will never be validated in the cloud is arguably (I'm trying to make the point) better than sending any wake words to the cloud. Less data in the cloud is better for user privacy.
This is what I mean by "blind spot", Amazon is not validating wake words in the cloud - it's filtering out REAL wake words that it thinks might be TV words.
Siri on the other hand does no such filtering, so is arguably worse. You think that's it's better, but that's because you have not thought this all the way through.
Will think on the rest, and do some more reading. I appreciate the responses.