I remember, in 2006, doing a full multimodal (touch, mouse and voice) web application with W3C standards: VoiceXML, XHTML and a glue dialect called X+V (meaning "xhtml plus voicexml"). The browser at the time was Opera, and the voice recognition was handled client side, with a free IBM ViaVoive plugin that Opera could download with a single click. It already had many things built-in to sync voice events to DOM events.
It's nice to finally see it embedded in major browsers.
It is not clear to me, however, whether it works offline or uses a web service.