2 Comments

1. 

   Dictator - Feb 13, 2009

   I just saw a press release that said that Nuance sold 10,000 licenses to the military.

   So, I guess they don’t think of this as research, and I guess that the number is now 90,000 doctors using it.

2. 

   Jerry Colburn - Mar 3, 2009

   I note that Nuance, owner of NaturallySpeaking, sponsors your blog.

   “Owner” does not mean “maker.”

   In fact, the company did not even make up its own name, Nuance.

   Rather, the company originated as a mid-1990s Xerox spinoff, named ScanSoft, originally focused on marketing paperless office products.

   Jim and Janet Baker created NaturallySpeaking at their Dragon Systems company, funded largely by DARPA , the Defense Department’s advanced research projects agency.

   Jim and Janet met while doing their Ph.D.’s at Carnegie Mellon, his in math and computer science, hers in biology, with a dissertation on hearing in whales.

   Jim provided the esoteric math modeling that eventually became the basis for computers guessing what humans hear in speaking to one another.

   By 1990 Jim and Janet had married, started Dragon Systems, and begun selling their first speech recognition product, DragonSpeak.

   This created a highly portentous microclimate within the 1990s tech bubble atmosphere.

   Usable speech recognition always seemed within reach. Consequently, the technology attracted a following, including IBM, which put itself in competition with Dragon Systems to produce the first “continuous” — as opposed to word by word — mass-market product.

   At the same time, in Belgian an accountant and a salesman created their own speech recognition company, named after themselves Lernout & Hauspie.

   Yet, usable accuracy remained almost within reach, always almost.

   The eventual crash saw the Baker’s sell Dragon Systems to Lernout & Hauspie for $350 million in stock, which in turn became worthless following a major August 2000 book-cooking scandal that sent the two Belgians to jail. IBM effectively wrote down its own speech recognition effort, turning the product over to Lernout & Hauspie, which ScanSoft acquired in a November 2001 bankruptcy auction in a New York courtroom, for $10 million cash and some stock. Oddly, the Wisconsin state employees retirement fund provided most of the cash.

   ScanSoft developed a mergers and acquisition business plan that led to its buying SpeechWorks and Nuance, both in the call center automation market. The company spent almost nothing on research and development for NaturallySpeaking, and the software’s post-Baker releases still use the speech engine from the Baker’s 1999 release 4.0.

   NaturallySpeaking works well for users who have dictionary pronunciations of the words in its large vocabulary. ScanSoft/Nuance has tried to improve accuracy by larding the program with “natural language” technology acquired along with SpeechWorks and that Nuance. This is simply makes phrases part of the vocabulary. It does not address the phonetic issues that become speech recognition language barriers for users who speak either dialects or English as a second language.

   You cannot have standardized deployments in healthcare unless you have high accuracy and usability for all users. Our physician and nurse corps have a high percentage of dialect and English as a second language speakers. You have to provide for them if you want national standards for electronic health records. Only speech provides fast enough input for practical medical reporting by physicians. You have to start with input, not databases to have a system that works. Small wonder the legislators have not checked to be sure they have included all the zeros in the 20 billion-dollar appropriation for a system that still lacks pilot deployments proving that it will work.