Item — HN Reader

You've got it backwards. Modern speech recognizers have a vocabulary of a million words and multi-gigabyte models. It's generally much more accurate to do speech recognition in the cloud, since you have more processing power and more RAM to hold large statistical models.

The rumor is that Apple is sending the audio to Nuance servers, i.e., they're doing cloud-based speech recognition.

2 comments

False, they're doing on device speech recognition. Servers are only used if you request information from the Internet. Look at the demos and read the hands on reports. On device recognition makes it much more usable than Google Voice Actions.

+1 on the cloud advantage.

I tried Google Voice Actions on my Nexus One quite a while ago, but it was optimised for the US market. The accuracy for me was so bad that I didn't bother with it.

Then recently, a Google blog on RSS said the latest app had been optimised for my locale. Now, of course, it's spookily accurate.

You can't beat your algorithms in the cloud being bombarded with sample data round the clock.