BestMatch

Last Modified: March 22, 2000

(Editor's note: This topic was originally written in June of 1998. With version 4.0 of Dragon NaturallySpeaking, Dragon introduced the new BestMatch III models which are even better than the BestMatch models described here (see NatSpeak Version 4.0 for more information). BestMatch models are still available in Dragon NaturallySpeaking version 4.0 for slower processors are systems with less memory.)

The most significant improvement or new feature of Dragon NaturallySpeaking version 3.0 is the improved accuracy. To get better accuracy, Dragon Systems has made a number of changes to the speech recognition system and data files. Collectively these changes have been given the name "BestMatch".

BestMatch is just a name. I think that it allows Dragon Systems to say that "Dragon NaturallySpeaking has BestMatch technology" and none of the other competing products do. The bottom line is that version 3 of Dragon NaturallySpeaking is more accurate.

Here is a brief list of some of the changes which were made in version 3.0 which contribute to the improved accuracy.

-- Dragon NaturallySpeaking version 3.0 has new acoustic models which do a better job of matching a wider variety of speech to phonemes and words. These new acoustic models are larger than the acoustic models which were used in Dragon NaturallySpeaking version 2.0. Because of this, using these new larger models is optional. I believe that switching to the version 3.0 (BestMatch) acoustic models requires an extra 4 to 5 MB of RAM.

-- Dragon NaturallySpeaking version 3.0 has new vocabulary files which do a better job at predicting word usage. The new vocabulary files in the Standard Edition and the Preferred Edition both have more words and more statistical information. These new BestMatch vocabulary files actually have over 45,000 words instead of the standard 30,000 words. In the Professional Edition of Dragon NaturallySpeaking version 3.0, you also have the choice of using an even larger vocabulary file which has over 64,000 words. The 64,000 word vocabulary also has even more statistical information relative to the number of words than 45,000 word vocabulary included in the Standard Edition and the Preferred Edition.

However, the larger vocabulary files also require additional memory. I would assume that the BestMatch vocabulary files in the Standard Edition and the Preferred Edition would require approximately 8 MB of additional RAM. And I would allow for 16 MB of additional RAM when using the 64,000 word vocabulary in the Professional Edition.

-- The vocabulary files in Dragon NaturallySpeaking version 3.0 now contain a significant number of trigrams. Trigrams are the name of statistical data that models the probability of words in sets of three. The addition of trigrams makes Dragon NaturallySpeaking that much more accurate in the presence of homonyms.

-- In Dragon NaturallySpeaking version 3.0, the command versus dictation discrimination logic has been completely rewritten. This means that Dragon NaturallySpeaking version 3.0 is much less likely to mistake a command for dictation or vice versa. (However, it also means that if you speak a command which does not exactly match the grammar, it is more likely to be typed.)

-- Also in Dragon NaturallySpeaking version 3.0, the selection logic was reworked to be faster and more accurate. This is the logic which allows you to say "Select " or "Correct ".

-- Dragon NaturallySpeaking version 3.0 also has the capability of using significantly more processing power to increase the accuracy further. This is optional, but if you have a very fast processor, you can change the tradeoff between speed and accuracy to take advantage of your additional resources.

-- Finally, Dragon NaturallySpeaking version 3.0 includes some other algorithmic improvements which have small but measurable impact on the accuracy.

In summary, BestMatch is the name of the improvements in Dragon NaturallySpeaking version 3.0 which improve your accuracy. Dragon Systems internal tests indicate that when using the larger acoustic models and larger vocabularies along with the other improvements in version 3.0, users experience approximately 25 percent fewer errors. However, the experience of heavy users is more pronounced. Everybody that uses Dragon NaturallySpeaking version 3.0 with BestMatch technology notices a significant improvement in the accuracy.

On the other hand, there is a downside to the BestMatch technology. Using the larger acoustic files and larger vocabulary files requires both more RAM and more processor. When using these improvements, you will need at least a Pentium 200 MHz processor and 64 MB of RAM instead of a 133 MHz processor and 32 MB of RAM. However, you can decide to use the older, less accurate but smaller, acoustic and vocabulary files when creating a user if you do not have a powerful enough system. You will still get the benefits of the algorithmic improvements but you will not see as pronounced an improvement in the accuracy.

BestMatch may be a buzzword, but Dragon NaturallySpeaking version 3.0 is a lot more accurate than previous versions and you will never want to go back.

By the way, you will have to rerun General Training if you want to take advantage of the BestMatch acoustic models. But if you're using Dragon NaturallySpeaking version 3.0 without upgrading to the larger acoustic models, then you will not have to retrain.

This web page (http://www.synapseadaptive.com/joel/BestMatch.html) was last modified on March 22, 2000. The contents of this page are (c) Copyright 1998-1999 by Joel Gould. All Rights Reserved. See Copyright Information for more details.