Google Voice Search
Google released an application that sent recorded speech to a server and returned search results, with no training on the individual user.
Why it matters
Speech recognition became free and ubiquitous, and its quality now depended on how many recordings had been collected.
The acoustic model was assembled over three years from the GOOG-411 directory service, which Google ran precisely for the voice samples. The computation happened on the server, so the phone could be weak. That closed a loop: more users give more data, more data gives a better model. Three years later acoustic models began to be replaced by deep networks and quality jumped again.