How to achieve great accuracy

How to achieve great accuracy with Dragon

Dragon Accuracy

Under optimal conditions users should achieve between 98% and 99.5% accuracy with Dragon Professional Individual (DPI) 15 out of the box

There are a number of variables that will help you achieve great recognition accuracy with Dragon:

  1. As software, Dragon works the way it's designed to work. Generally speaking, software only represents about 5 to 10% impact on overall accuracy. Installed and configured correctly lays the groundwork. From that point going forward accuracy is primarily a combination of the users dictation style and the quality of the microphone/soundcard being used. By quality we don't mean $$$. What we do mean is that the microphone and soundcard being used must output a clear and clean audio signal. Electromagnetic interference (EMI) and/or poor connections between your microphone and/or your soundcard, high-volume background noise, persons talking in the background, incorrect positioning of the microphone (element), speaking too softly or too loud, over articulating (exaggerating) your pronunciation of words, and attempting to dictate single words or to dictate in short choppy phrases of three words or less are the most common problems relative to failure to achieve good accuracy, but there are others as well.

This page is an excerpt from VoiceComputer's accuracy lesson.

VoiceComputer features an interactive 'Virtual Dragon Trainer'. - Help and training that you can access with simple voice commands as you are working

VoiceComputer's help and training will help you guarantee your success with Dragon with dozens of lessons, demonstrations, tutorials on Dragon and hundreds of 'cheat sheets'.

Achieving great recognition accuracy with Dragon (continued)

  1. Processor (CPU) type and speed, CPU cache (L1, L2, and/or L3) and RAM can affect the overall performance (i.e., time from dictation to display of text) of Dragon, but they don't have any impact on accuracy.

  1. 70% of accuracy success or failure is user dependent. What that means is how a user dictates has the greatest impact on achieving great recognition accuracy with Dragon.
  • Clearly enunciate words.
  • Dictate at a normal pace at a normal volume using a normal tone of voice. Don't speak too loud and don't dictate in such a sloppy manner that you run your words together.
  • Use proper correction techniques. That is, correct full phrases (utterances) containing misrecognitions vs. correcting words. Correcting individual words will not improve your accuracy. Dragon's Active Vocabulary sorts entries in alphanumeric format: blank Written form's first, numerical entries second, and alphabetical entries third, such that words that are pronounced the same (homophones) or which have pronunciations that are close to one another can be misrecognized because Dragon will generally choose the first alphanumeric entry that it finds when searching the vocabulary. Words such as "and", "in", "end", "weight", "wait", "eight" and "ate" will be chosen by Dragon based on the alphanumeric order in the Active Vocabulary (Vocabulary Editor). These types of words require context that Dragon understands. Remember, Dragon does not understand the meaning of words, nor does it understand what you meant to say. Dragon is simply not that capable (smart). This applies to all versions of Dragon, including DPI 15.
  • Try correcting on-the-fly. While you can wait until you finish dictating an entire document before making corrections, it is much more difficult to remember what you actually said the longer you go without making corrections. In addition, depending upon how you set up Dragon, the playback of your original speech may not be present and accessible when dictating large documents before making corrections. If you wish to wait before making corrections, do so in terms of paragraphs vs. entire documents. It is just as fast and more efficient to say "correct that" when you see the Dragon has misrecognized something that you said than to wait until you finish dictating entire documents.
  • DPI 15, unlike previous versions, automatically saves your user profile. Just as you will lose new words entered in the vocabulary if you bypass this feature, you will also lose any adaptation (learning). Actually, this is not unique to DPI 15. It has existed all along in all versions of Dragon. What is different is that the new speech engine in DPI 15 is more efficient.
  • One of the major reasons why users experience profile corruption is because they don’t properly correct misrecognitions in their documents before saving their user profile(s). Failure to properly correct, save, and close their documents actually teaches Dragon to continually make recognition errors, sometimes to a point where correcting such becomes next to impossible.

 

Achieving High Recognition accuracy with Dragon (continued)

  1. Many users make the mistake of assuming that speech recognition is like talking to another person. Nothing could be further from the truth. Even when two people are talking to one another it is possible to fail to understand what the other person is saying, or attempting to say. Consider the following example. If you and a friend are at the train station and your friend goes over to the snack bar, turns around to you, and puts up one finger and then two fingers. How do you know what he or she is trying to communicate. He or she could be saying that it will take one or two minutes before they get what they want, or they could be asking you whether you want one or two of whatever they’re getting. In short there could be a myriad of reasons behind what they’re attempting to communicate. Without any underlying context, the probability of your understanding correctly what they’re trying to communicate becomes exponentially more complex and probabilistically less likely. This is the problem that speech recognition runs into in recognizing what you are trying to say.

On the other hand, if your friend asks you before going to the snack bar if you want a hamburger and then puts up one finger followed by two fingers, the probability of understanding what they mean is much higher (better). In terms of how Dragon functions, when you dictate in longer phrases, sentences or paragraphs, the probability of Dragon understanding what you said and what you meant to say is much higher (better). Dragon uses what are called n-gram models (i.e., monogram, bigram, triagram and quadgram). These are the context models that Dragon uses to analyze the most common usage of words in your dictation. For example, consider the following utterance:

 

Looking confused, Jack came to the end of the road and then turned right.

Dictated as one continuous utterance, Dragon is able to predict the proper usage of both “and” and “end” in this context. However, without this context it is much more probable that Dragon will get the difference between these two words wrong, and the probability of such increases the shorter the utterance or relative to dictating each word by itself.

Understanding and using these techniques will improve your accuracy over time. Remember, just as you are in control of your own destiny, you are also primarily responsible for your own accuracy when using speech recognition.

Chuck Runquist

chuck@voiceteach.com

Be pain-free at your PC

Computer Control, by Voice!

The idea behind VoiceComputer and Intag was to build the ultimate desktop experience. We re-invented/re-engineered computer control to take the repetitive out of repetitive stress, the stress out of repetitive computing.

Now you use voice control instead of your mouse or touchpad or use your voice in conjunction with your mouse/touchpad.

We make accessibility simple and productive.