Speech Technology Magazine Blog » Nuance DID Have to Train Vocalizer

In response to my post yesterday about Nuance’s Vocalizer competing with humans in the Boston area to see which could correctly pronounce several rather difficult names and phrases, I asked for clarification as to how much work had to go into the Vocalizer solution to make it fare so well. This is what I got back from Nuance:

Nuance recently updated Vocalizer to support more than 150,000 first names and 1 million last names, based on U.S. Census data and other sources.

o   This covers about 95 percent of the U.S. population – so Vocalizer gets names right more than 95 percent of the time (and first names more than 99 percent of the time!)
o   Likewise, Nuance is now accurate more than 98 percent of the time for city names, including even obscure places like Natchitoches, Louisiana; Kalalach, Washington; and Weippe, Idaho.   (That’s “NAK-ih-tish”, “KLAY-lock,” and “WEE-ipe.”)

-          Because new words are always being invented (w00t, locavore, tweetup, freemium—and who could forget, refudiate!!!!), and because rare words can suddenly become frequent (e.g., vuvuzela, the South African trumpet heard incessantly at the 2010 FIFA World Cup), Nuance also publishes regular updates supplementing its dictionary for customers to use.

Well there you have it. So I congratulate Nuance on a fine product, and a fine showing for speech technologies in general.

via speechtechblog.com