Hacker News Reader

Comment by sneak | Hacker News Reader

The speech recognition is performed in the cloud, not locally, so under the current architecture they do indeed need to upload the audio capture following the wakeword.

One of the other things I used them for frequently before disabling it on all my devices and removing all of the networked mics from my house was things like "remind me to $THING at 21 tonight".

View on HN

Fwiw, starting in iOS 15, Siri will process speech on-device, which is huge.

It's other on-device processing that's planned in iOS 15 (such as scanning my local files without consent to report on me to the police) which is going to keep me from using devices with it.

I hope this doesn't make Siri's quality even worse, but it seems inevitable.

It's much slower even with the A12's dedicated cores for this purpose, but I assume that'll improve over time.

Try this - turn on airplane mode and go into the notes app and do a dictation.

On the iPhone, the speech to text can be done on the phone. I'm not sure if this is the case for Siri, but it doesn't need to be done in the cloud.

That said, the words to intent mapping doesn't appear to be something that the on phone model is able to handle.

> The speech recognition is performed in the cloud, not locally, so under the current architecture they do indeed need to upload the audio capture following the wakeword.

I bet it wouldn't have to be if its vocabulary were restricted to timer related words.

It doesn't have to be, but it makes the device cheaper and they're going to gobble your data up in the cloud anyway.

Reply on news.ycombinator.com