Speech Recognition - Ready for Prime Time?

The Contenders

When I first made the decision to try out speech recognition, there was an overwhelming favorite on the market: Dragon NaturallySpeaking. I had never used it before, but I'd heard about it and it was generally well-regarded. I picked up a copy of Dragon NaturallySpeaking 8.0 Preferred and commenced using it. The training process took about 20 minutes, another 20 or 30 minutes was spent scanning my documents for words and speech patterns, and then it was basically done and I was ready to start dictating. I've now been using Dragon NaturallySpeaking for several months, and during that time training has further improved the accuracy.

Dragon isn't a particularly cheap piece of software, but when you consider the versatility it offers and the fact that I've already spent about $700 on a desk, chair, and keyboard in an attempt to make an "ergonomic workspace," spending another $100-$200 is hardly a concern. The $100 Standard version apparently has reduced functionality, though apparently the only major difference is that it lacks the ability to transcribe recordings. For home use and personal use, you can get a discount on the Preferred version and buy it for $160. Unless that extra $60 is really important to you, I would have to recommend going with the Preferred version -- you never know when the ability to transcrive a recording will come in handy.

Of course, Microsoft Office 2003 also has built-in speech recognition. I have never heard anyone really talk about it, and I have never tried it myself, but having become familiar with Dragon NaturallySpeaking I figured it was only fair that I give Microsoft's product a shot. After all, practically every business in the world has a copy of Microsoft Office 2003 installed, so perhaps there isn't even a need to go out and purchase separate speech recognition software. One other item that may be of interest is how much processing time each product needs. Voice recognition may or may not benefit from dual core processors, but there's only one way to find out.

I conducted testing on several systems, but eventually settled on using one for the actual benchmarking. If there's interest, I can go back and look at performance on other systems, but for the most part I have found that modern Pentium/Athlon systems are sufficient - with a few exceptions that I'll get to in a moment.


Test System:
AMD Athlon X2 3800+ @ 2.60 GHz (10x260HTT)
2x1024MB Patriot SBLK @ DDR-433 (CPU/12)
Western Digital 250GB 16MB SATA-2 HDD


I began using Dragon NaturallySpeaking on a single core Athlon 64 3200+ socket 754 Newcastle (@2.42 GHz) -- my old primary system, which I have been using for about 18 months. I finally broke down recently and decided it was time to move on to a dual core setup for my main system. Both systems are of course overclocked, because that's the type of user I am. Since this is a look at a software technology as opposed to a hardware article, the system clock speed isn't particularly relevant except as a guideline of what level of performance you can expect.

The major reason for the upgrade is gaming - the old AGP 6800GT wasn't cutting it anymore, and the only reasonable upgrade required PCI Express. (That should tell you something about the amount of processing power most business tasks require - the 754 platform is still more than sufficient for most people!) I figured since I was already switching to socket 939, there was no reason not to add a second processor core. That extra core does help out when I'm trying to do multiple things at once, and Dragon does tend to consume a decent amount of resources. MMO gamers might find it useful as a way of chatting without having to type (and it might just cut down on the use of annoying abbreviations if more people did it, but I digress...). When I'm only dictating, though, I don't really notice the difference between my old system and my new system as far as speech recognition is concerned.

So how do you test and benchmark speech recognition packages? The more real world a test is the better, and what could be more real world than an article written for our web site? How about this very article? I'm going to take the first two pages of the article in their present form (minus the Isaac Asimov quote and potentially some later edits) and dictate the text into a sound file. All punctuation will be dictated, and I will edit the final sound file to remove any speech errors. The final sound file will be played back for both speech recognition packages, and with 1181 words of text we can come up with an accuracy rating.

This first sound file is basically my "dictation voice". There are two elements to training a speech recognition program: first, it learns to recognize your voice; second, you learn to adapt your voice to improve accuracy. After creating this first sound file, I realized that my voice didn't sound very normal to me. I'm okay with that, but I decided a second sound file was needed to stress test the software packages. I read the text a second time for this sound file, with a few minor updates to the text, but this time I spoke in a more natural voice and I didn't go back to correct any errors. I won't count any of my errors against the accuracy score, but this will hopefully provide additional insight into how these two voice recognition packages perform.