How do I improve my computer's voice recognition?

March 3, 2009 10:01 AM

How can I improve the accuracy of Dragon Naturally Speaking 10 or some other voice recognition software?

I'm attempting (and having some luck) overcoming writer's block with Dragon Naturally Speaking 10, but I'm not getting anywhere near the accuracy levels people rave about. Since I can rarely get a full sentence to be read in accurately, I'm guessing that I am doing or using something that makes dictation a lot harder than it needs to be. There's got be a way to get better than the pretty dismal 80% accuracy I'm working with now.

I'm running Dragon on a couple year old Thinkpad T60 with XP and 2 gigs of RAM. I've used a Koss SB45 microphone that plugs into the laptop's sound card and a Plantronics MX500i that uses a 2.5mm port and USB adapter. I have also done a few of the training sessions, though they haven't seemed to help.

Does anybody have any equipment recommendations or training tips that could help me use voice recognition software?

posted by eisenkr to computers & internet (5 comments total)

The only recommendation I have is to do all of the training sessions. Training the program will tune it to your inflection and style of speech. That's the reason there are so many of them. Unfortunately, voice recognition is an imperfect science, and Naturally Speaking is the best out there. The program necessitates slow, clear speech at a speed slower than a normal conversation.
posted by orville sash at 10:05 AM on March 3

I would try using the "Play that back" feature just to make sure your voice is clear.

Here are several things affecting Dragon that have come up in my office:
-bad microphone/connection, sounds like lots of static
-laptop not switching from built in mic to the mic you just plugged in
-not speaking naturally, instead of using a natural tone, the user would talk slow and deliberately.
-incompatibility with text editor program, we have a proprietary text editor and for some reason or another Dragon likes to do stupid things while dictating into it but works perfectly in MS Word.

We have had great results with the Andrea Electronics NC-181 mic. Its actually the mic thats bundled in the Dragon retail box.

The program necessitates slow, clear speech at a speed slower than a normal conversation.

This was the case with the older versions of Dragon, but version 10 is really good at recognizing conversation paced speech. When our office transitioned from 9 to 10, it was an amazing difference in recognition speed and accuracy.
posted by aGee at 10:35 AM on March 3

Seconding orville. When I used it, it took doing all the exercises and also persistantly correcting it, and it gradually learned. It's been years since I did use it (version 7), but I think there was one way of correcting it's errors that would lead it to learn from its mistakes and one that wouldn't (but was quicker). Highlighting and letting it try to find an alternate before typing it in as opposed to just typing over the mistake, or something like that.

Also, you may need to work on speaking style. You have to be clear and consistant, like you're a newsreader on TV.
posted by K.P. at 10:41 AM on March 3

Speaking as someone who's worked in speech recognition, 80% isn't dismal - it's fantastic.

You can get better by training the program as best you can to your particular vocal idiosyncrasies. You'll never get to 100%. You might also get incremental improvements w/better audio hardware, as someone else already mentioned.

Otherwise, speak slowly, clearly, and try and suppress whatever accent you might be bringing to the table.
posted by NoRelationToLea at 11:19 AM on March 3

While DNS 10 itself may not necessitate speaking more slowly, hardware on the lower end of the requirements scale could be the cause.

Reduce background noise in your environment. No music, TV, etc.

Make sure that the mic is the proper distance away from your mouth and in the correct location.

Speak clearly, don't mumble. Someone above had it right; the more you speak like a newscaster, the better (good articulation, volume control, even pacing).

Also, while the software will adapt to you, you will also adapt to it. You will get better results as you go along. It's important to do the corrections by voice and not by mouse or keyboard. DNS will not learn if you correct manually.
posted by reddot at 8:41 AM on March 8