What a hip granny you are! There ought to be a way to make this a "sticky" post so that people can find it without much effort.

Bruce

This explanation is OK, but it is not technically correct. The process is more complex than this and involves a step by step conversion of the analog input to digital; comparison to the Acoustic Model to produce a binary phonetic equivalent of words, phrases, and pauses based on the speaker's pronounciations (training & Vocabulary) and the settings in the DNS Options dialog for pause timing (i.e., how pauses are detected); the results are then analyzed using the Language Model for context. The language model consists of a complex set of Hidden Markov Models based on the precalculated probability coeficients derived from common usage. There are several models and a number of calculation algorithms used to do this, but essentially each branch of the bigram/trigram HMM trees are searched for the most likely match. Whether a bigram or trigram model is used is based on the length of each utterance preceding a pause. If an utterance consists of less than 5 words, the bigram model is used. If 5 words or more precede a pause, the trigram model is used. These models work by analyzing the right and left contexts for each three word (bigram) or five word (trigram) combination. Next the 9 best matches in order of probability are calculated and the match with the highest probability is selected for display (transcription).

Therefore:

1. What you train and how you train affects accuracy.

Training individual words, even in the Vocabulary Editor, has no impact on accuracy because all the words contained in the active vocabulary (160,000 words loaded at runtime) and/or the Background vocabulary (what is stored in C:\Documents and Settings\All Users\Application Data\ScanSoft\NaturallySpeaking8\Data\Enx\enx_aus_general_large) have a fixed probability that does not change regardless of training. Changing the probability coeficeints of these words would result in accuracy degredation on a scale that would render DNS useless. Suffice it to say that this is necessary for the proper recognition of words. Remember that the training script makes reference to "...words by themselves AND in the context of other words." What this means is that there is a distinction between the handling of individual words that can be modified only by training words in context.

Training words in context alters the context probability coeficients because this is where DNS distinguishes individual words by the context in which they are used, and this is what the AND means in the script phrase above. So if I just say "an" and get "and" it is because in the vocabulary and in common speech "and" usually occurs before "an". Training these two words individualy will not change this probability. But, traing the phrase "He had an awful cold." will increase the probability that "an" will be correctly used in this context. Therefore, this is why analyzing documents increases the accuracy of words used in context based on how the user writes/dictates.

It is important to remember that how we say words and phrases during dictation is unique and distinctly different from how we say them when we train them. For example, when you dictate the phrase "Training individual words, even in the Vocabulary Editor, has no IMPACT on accuracy..." and it is transcribed as "Training individual words, even in the Vocabulary Editor, has no IMPORT on accuracy...", and you say "Correct has no IMPORT on accuracy", we all pronounce that phrase differently when taken out of the context of the entire sentence. In addition, we virtually never say the same thing in the same way twice anyway. Therefore, when oorrecting this phrase to change "IMPORT" to "IMPACT", if the playback in the Spell dialog is clear and correct, retraining the phrase alters the acoustic equivalent and, done too often, can result in eventual corruption of the words in that context. Multiply this by 100's of corrections per day and it is easy to see why some user files eventually become significantly corrupted.

On the other hand, if there is no playback, the user has to make a choice. Did they say it correctly and clearly or not. If in doubt, then train it. No training results in the continued link between the original spoken utterance and the correction. If this is, or was, incorrect, then the correction will be associated with an incorrect acoustical utterance. Training in this context does not hurt because it will only show the one phrase, not the corrected and incorrect words or phrases, and training will create a correct, even if slightly different out of context pronunciation, acoustical/word or phrase link. In addition, no training will result in a correction that will not be saved for later use by the Acoustic Optimizer. Only words/phrases trained during correction or during additional training are recorded for update by the Acoustic Optimizer. Remember that the key word is Acoustic, not Language relative to the Acoustic Optimizer.

2. General training should be done once.

It is only necessary to perform the general training one time. If accuracy, which should be tested before performing any additional training, does not produce at least 97 to 99% accuracy, performing additional training is not likely to improve accuracy and, in many cases where the accuracy is below 90% after initial training, will likely degrade such. The reason for this is that the hardware and/or user speech are the primary cause. In short, check your pronunciation and listen to your self speak via playback (we sound significantly different when listening to a recording or our speech than we do otherwise), and check your hardware. In most cases it is the combination of microphone, system performance and soundcard/input device that are at fault. You should never have to perform more than two general trainings, and a second training is only viable on rare occasions when you have changed the way you speak or your hardware changes, AND when your initial training has produced high initial accuracy. Beyond this, proper correction and enunciation will produce the best results in improving accuracy, and/or in maintaining it.

Keep in mind that performing initial training several times before using DNS is a waist of time. The reason is that the general training does not set the Acoustic Model based on words, it listens for and sets the user's pronunciation for each of the 40 or so base phonems in the given DNS Language version. This is why the the content of the training script does not matter as long as it is takes approximately 5 minutes or more to read. The Adaption process then mathes these to the Active Vocabulary pronounciations to form the user's Acoustical Model (i.e., how the user pronounces the phonetic reprentations that make up each vocabulary word).

3. The original question.

First, be careful how you deal with punctuation that you want spelled out. Adding anything to the Vocabulary that conflicts with current word forms, including punctuation, can cause sproadic misrecognitions. This includes using the Alternate form or the spoken form.

Second, in most cases you should be able to get the desired result by analyzing documents for your writing style in which you use such phrases frequently. One of the more experienced users who monitors this group uses a technique that is very creative. He creates documents using his list of most frequently used/misrecognized words/phrases in the context of a written document and used many times, then repeatedly has DNS analyze that document.

Third, add commonly used phrases to the Vocabulary. For example if you want "At one point" to be displayed that way, try entering it first as At one point with no special spoken form or Alternate form. If that works, use it that way. If not, try entering it as it is generally recognized (i.e., Written form asAt1., the Spoken form as At one point, and the Alternate form as At one point. This should work, but the rule of thumb is add only what you need, do it the simplest way, and avoid conflict with other vocabulary entries.

Lastly, Spoken forms work best when they are entered as quasi phonetic or phonetic equivalents. For example, if you want AKG dislplayed, set AKG as the Written form, Ae-Kay-Gee and the Spoken form, and AKG as the Alternate form (just to be sure it gets displayed the way you want).

Another final rule of thumb. Don't blame DNS until or unless you are absolutely sure that the problem is not a result of PEBCAC (Problems Exists Between Computer and Chair). That is said tung-in-cheek and not intended to offend. It is a simple technical support joke that recognizes that most problems are solved at home.

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