Watch what you say about your computer: It's getting better at understanding your voice.
Nuance Communications Inc. says the latest version of its speech-recognition software can achieve -- with some speakers -- 99 percent accuracy out of the box, without a "training" session to familiarize the software with how a particular person talks. It was scheduled for launch Tuesday.
Dragon NaturallySpeaking 9 lets users control many applications on Microsoft Corp. Windows PCs and dictate documents within them. It's designed for professionals such as doctors and lawyers, as well as for enterprise employees who frequently have to type and for disabled people, said Matt Revis, director of product management for dictation solutions at Nuance.
The accuracy rate, or what percentage of words the software spells correctly by itself, varies depending on sound quality and how a person talks, Revis said. But Nuance has improved it by 80 percent since NaturallySpeaking 8 was introduced in 2004, according to the company.
Version 8 could reach 99 percent, but only after the user read a prepared script, Revis said. Now users can get that level of accuracy right after installing the software and starting it up, though a script is still available if a user isn't satisfied with the results on the first try. In any case, the software can continue learning on its own just through normal use, Revis added.
The out-of-the-box 99 percent accuracy figure is for the American English version, but the new Dragon releases for other languages get similar boosts in accuracy, Revis said. The software is available for Dutch, French, German, Italian, Japanese and Spanish, as well as for Australian, Asian, Indian and U.K. English. Also with NaturallySpeaking 9, Nuance is certifying two Bluetooth wireless headsets for use with the software.
The current version of NaturallySpeaking is "uncannily accurate," so the new version should work very well -- but only over time, in most cases, said Bill Meisel, president of TMA Associates, a speech industry consulting company in Tarzana, California. Most users are unlikely to see 99 percent accuracy the minute they start using the software, though after about two days of use their accuracy will improve, he said. In fact, Meisel questions Nuance's decision to emphasize immediate results. Many users quickly get frustrated, not realizing that it usually takes time to get good results.
"They're raising expectations higher than necessary," Meisel said.
Nuance has the market for dictation software pretty well sewn up outside of specialized areas, namely medicine, Meisel said. For one thing, it controls a formidable cache of patents, he said. However, voice recognition for controlling applications is drawing increasing interest. IBM Corp., Microsoft Corp. and Google Inc. are among the companies that may be working on bigger plays in this area, Meisel said. Microsoft already includes some voice-recognition capability in Windows.
NaturallySpeaking 9 is available immediately. The Standard version costs US$99 and the Preferred version, which includes support for more applications and for converting digital voice-recorder tracks into text, costs $199. A specialized version for the legal profession costs about $800, and iterations for various medical specialties cost $1,199 each, Revis said.
