I think you have it exactly backwards - the iPhone 4S/Siri speech-to-text/natural language processing are done in the CLOUD. The text-to-speech is done on the phone itself. My (non-Siri of course) iPhone 4's Voice Command stuff is COMPLETELY on the phone itself, and would do TtS of my contact list and Artist names, etc.

From the article:

> The iPhone 4S really sends raw audio data. It’s compressed using the Speex audio codec, which makes sense as it’s a codec specifically tailored for VoIP.

I don't know why this is being downvoted -- I guess others are reading something different than I am?

There are three parts to Siri:

1. Speech-to-text (parent has it backwards but that's what he means, obviously)

2. Text-to-intent (referred to by parent as NLP)

3. Intent-to-API calls

Obviously, (1) happens in the cloud and (3) happens on the device. It is still unclear where (2) happens but if the cloud service only responds with text, it seems that (2) happens on the device.

And (2) is still a hard problem by itself.