Stan,

First, most users tend to think of all of the various utilities in DNS as being unique and distinct from one another. This is not the case. All functions in DNS are related to the basic speech models: The Acoustic Model and the Language Model. The specific speech model to which a particular function, or utility, applies depends upon whether or not that function relates to the spoken form or the written form. For example, all training is linked to the Acoustic Model, and all functions that deal with writing style or context analysis are linked to the Language Model, of which there are 3: Vocabulary, bigram, and trigram models.

Second, the Acoustic Model and the various language models are linked via the "spoken form" and the "written form." The most common observation that you can make in this regard is in terms of how the Vocabulary is configured. The Vocabulary is the single most important Language Model. The reason for this is simply because without the Vocabulary, there is no link between the Acoustic Model and the various Language Models. It is via the Vocabulary that the written and spoken forms are linked. In the Vocabulary Editor, this relationship is most obvious because you have the option to enter a written form and a spoken form. If you leave the spoken form blank, than the Vocabulary Editor automatically assumes that the normal phonetic pronunciation of what ever word or phrase you are adding to the vocabulary is the default form for dictating that word or phrase. If you add a spoken form, then DNS assumes that the pronunciation of that word or phrase is based on the spoken form, and the phonetic equivalents for that spoken form are added along with your vocabulary addition. If you train a word or phrase added to the vocabulary that contains a spoken form, it is the spoken form that you are training and that spoken form is then added to the vocabulary linked to that word or phrase using the phonetic equivalent of your pronunciation. My point here is that this is how the Acoustic Model and the Language Models are related to one another. That is, through the link between the written form and the spoken form. So the basics of speech recognition in DNS, as well as in all the other SR products, is the relationship between the Acoustic Model and the Language Models, which are basically referred to technically as "speech models." This relationship underlies all of the utilities and functions in DNS.

Third, DNS does not learn from your dictation. It only learns via correction. In other words, this is a simple algorithm that says, in essence, “if what I say is not corrected, then it is correct as dictated by default.” If, on the other hand, what is dictated is perceived by the user as a misrecognition, then any time the correction dialog, whether that be the Correction Window or the Spell dialog, is invoked, Dragon NaturallySpeaking® assumes automatically that any changes made to the original dictation in either of these dialogs is, by default, a correction. Juxtaposed to this, if you select a word or phrase and redictate it, Dragon NaturallySpeaking® assumes, by default, that this is a substitution, not a correction, and therefore simply assumes that, for lack of a better analogy, you simply erased what you dictated originally and redictate it, or changed it.

Fourth, corrections are stored in both your user, unless you do not save your user when closing your DNS session, and in the appropriate inclusions relative to the Acoustic Optimizer. When you run the Acoustic Optimizer, the Acoustic Model, along with the appropriate links to the various Language Models, is reanalyzed based on all of your corrections made in between the current running of the Acoustic Optimizer and the last running of the Acoustic Optimizer. In DNS 8 your entire user is a readapted using the same algorithm that displays the “Adapting User” at the end of general training. One of the reasons why, in DNS 8, the Acoustic Optimizer displays the time in terms of several hours is that it is going back through your entire acoustic profile and checking it against all of the corrections that you have made since the last time that you ran the Acoustic Optimizer. It then readapts your Acoustic Model based on these changes. In short, it's not dissimilar to completely re-adapting your user. In fact that is exactly what's happening, even though it only takes between 10 and 30 minutes of actual time to do so.

Lastly, what happens in the correction dialogs is this. If you "choose" an option, the original “spoken form,” what was recorded by DNS when you originally dictated the word or phrase that you are correcting, becomes linked with the written form that you "chose." This is why it's so important to have the automatic playback on correction enabled. This allows you to listen to what you said when you originally dictated what you are correcting so that you can make a decision as to whether to keep the original "spoken form" or retrain it. Since we've already covered training we won't go into that. However, if you choose to train a correction, then the New "spoken form" is then linked to the New “written form.” On the other hand, if there is no playback when you select a word or phrase for correction, then whatever is associated (i.e. spoken form) from any point in the past where the original word or phrase was dictated and recorded in the Acoustic Model is retained as the spoken form associated with your "correction."

The basic answer to your question, after the above diatribe, is simply this. DNS only learns from your corrections, DNS assumes that a correction is being made only if the correction function is used. Any other form of replacing text in a document that does not invoke the correction function, is not a correction, but rather a substitution. DNS treats substitutions as if they were a normal part of your dictation. That is, if a substitution is not corrected, then it is assumed to be correct as dictated in terms of the link between both the spoken form and the written form.

Chuck Runquist
Former DNS SDK & Senior Technical Solutions PM for DNS with Lernout & Hauspie (L&H)