Review: Speaking With Dragon Is Ever More Natural

Nuance Friday released Dragon NaturallySpeaking version 10. Yes, the same bells and whistles are still there from previous versions. Yet, key improvements in the new speech engine make the software run even faster. Delays when transcribing were hardly noticeable. But what makes version 10 far more interesting than previous iterations is a new acoustic technique that adapts to regional accents, even foreign ones.

NaturallySpeaking comes with a new option that lets you select accents from different regions in the U.S., including British and Australian accents, Indian and Spanish accented English. Nuance is so confident about the accuracy of NaturallySpeaking when working with accents that training is now optional.

However, we found that the best way to acclimate NaturallySpeaking to your voice is to go through the short training session, which takes about 10 minutes. After initial tests transcribing our dictation with Microsoft Word, we found it remarkably fast and accurate. The software didn't skip a sound. The buffering worked well on a 1GHz notebook with 1 GB of memory.

To achieve close to realtime transcription, Nuance invented a unique statistical model that analyzes words in a sentence. The software chooses the most likely words to add into it based on the proximity of other words. Essentially, the selection process picks out words that are likely to follow other words in phrases. But instead of using a library to match words, NaturallySpeaking does something far more interesting. The model relies on sounds rather than words. The software analyzes what sounds are likely to follow other sounds in phrases. Based on its sound scanning, NaturallySpeaking chooses the likely words. This is why version 10 is blazing fast.

NaturallySpeaking is now Web friendly—very Web friendly. With Internet Explorer, you can navigate the Web by using the Google and Yahoo search engines. You can even do a little research in Wikipedia and shop at Amazon. The navigation technology is quite intuitive but quirky. On some tests, we were not able to get NaturallySpeaking to follow directions. We ended up using the mouse to do the same task a few times. We needed to switch between programs to keep the right ones in scope. Moreover, we found it tough to jump between programs. The software delayed reacting to new verbal commands. Although this product has come a long way, it still has a long way to go to make Web navigation seamless.

Programs like Microsoft Word are far simpler, so the commands are more responsive and easier to learn. Nonetheless, Nuance added some user-friendly features to improve navigation. Just by saying numbers, the software adds numbers next to links and Web elements on a page. Adding numbers to a Web page helps NaturallySpeaking pinpoint names of links. Other commands like "click text link" are simple to remember and use.

You can also use NaturallySpeaking to transcribe recordings of phone conversations or meetings. Creating a new acoustic model for phone voices doesn't take too much time, except that it has to be done with at least one other person speaking over a phone. If the signal is clear, the training process is quick and painless. The software is sensitive enough to pick up sound discrepancies and volume changes over phone lines.

Verbal commands are now more accurate, as long as you don't switch between programs quickly. With Word, you can edit copy by mixing verbal commands and normal speech in realtime. We dictated a few paragraphs and mixed them constantly with verbal commands. The software performed extremely well.

For around $99 you can get a standard version of NaturallySpeaking. Nuance also sells a Preferred version with a Bluetooth headphone for around $199. Is it worth it? We think so.