Available for two years, the iPhone and Android versions of Google Voice Search do a respectable job recognizing speech on mobile devices -- it even recognizes Chinese -- but it's not perfect. Each person speaks with such different speed, volume and inflection, accuracy is a challenge, especially when using a single voice model for all speakers.
Google says this new version does a better job of solving that problem. According to the official Google Mobile Blog, "Although subtle, accuracy improvements begin fairly quickly and will build over time." Voice Search accomplishes that feat by recording each word its user speaks, then matching that speech with words it correctly recognizes, gradually learning from its mistakes.
Google satisfies privacy concerns by making it easy to turn off this recording capability.