Not exactly correct. If the playback is clearly what you said, or intended to say, and the correct word or phrase is in the selections shown, then select the correct word or phrase, but don't train it. Again, the reason is that what you said originally was said in the context of your total dictation. If you train your correction under this condition you run the risk of training the word or phrase out of the total context of you original dication (i.e., we tend to speak differently during normal speaking than when we pronounce a word or phrase during correction). This changes the Acoustic Model representation for the way you normally speak. If the difference is significant, then you alter the spoken form such that it may result in increasing the propensity for the correction to be misrecognized. It is important to note that a one time correction/training will not generally result in future misrecognitions. However, doing this continually eventually alters the Acoutic Model in a negative way, thus causing accuracy degredation, which requires correction, which further exacerbates the problem. In short, it is the repitition of continually correcting and training the same word or phrase over time that is one of the causes of accuracy degredation.

The other part of this is that if a word or phrase is misrecognized, you should always correct it, but only train it under the above conditions. DNS does not learn from your normal dictation. The speech models and the recognizer work on the principle that if you do not make a correction, what was said and what was displayed are correctly associated. Only corrections train the speech models, and DNS only learns from correcting misrecognitions. Acoustic and Language models are distinctly different and how each is modified is distinctly different. The Language Model is only modified by document analysis. The Acoustic Model is only modified by training. When performing corrections, selecting the correct word or phrase from the selection options, or typing the correction, is based on the document analysis algorithm and changes the Language Model (context). Training changes the Acoustic Model. Performing corrections links the Acoustic Model to the Language Model either in terms of what was said originally or in terms of any training done during correction. It is critically important to remember these distinctions. This is the way SR works. Unfortunately, it is not subject to debate. SR algorithms work the way they do whether we disagree with them or not.

First, if what you mean is that Automatic playback on correction is turned on, but you do not always have playback when you correct something, then your initial statement is valid. However, not always wanting to take the time to follow this procedure is ignoring how DNS works. Since you take the time to correct a misrecognition, it does not add significantly to the time required to ensure that you use the correction function correctly. Not doing so is anathema to improving accuracy. There are three principles involved here.

1. Practicing correct correction techniques saves time because it increases overall accuracy and reduces the necessity to continually deal with misrecognitions.

2. Practicing correct correction techniques continuously results in their becoming instinctive and automatic, which in turn saves time.

3. Failure to practice correct correction techniques generally adds 3 to 4 times the amount of time required in overall loss of dictation speed and general loss of productivity. If you save 10 minutes by cutting corners you add up to 40 minutes to your overall time to complete dictation tasks, and loss of productivity increases the more you practice incorrect techniques.

Second, if your user accuracy is optimal, then the correct selection option in the Spell dialog almost always shows up as #2. Occasionally it will show up as #3 or #4. Rarely, if ever, will it show up as #5 or greater (keep in mind that DNS only selects the 9 highest probability matches whether in the correction function or during normal dictation). If you are getting the misrecognitions you cite as "wacky" phrases, then something is very wrong with either your hardware, your user, or some combination of the two.

Third, correct phrases containing misrecognitions, not words. This adjusts (corrects) the context recognition (Language Model) exactly the same way that analysing documents does. In fact, both of these functions use the same algorithm. Correcting individual words simply does not work unless the word being corrected is "wacky" (i.e., totally wrong and there is no selection in the Spell dialog or Correction window. Whether this occurs with regards to words or phrases, you should train both the correction and the misrecognition. This is probably the one circumstance where training should always be done. Otherwise, train words by using either the Vocabulary Editor or the "Train a word" command. Vocabulary probability (words by themselves and not based on context) is fixed and never changes regardless of training or correction. The only changes that are ever made to the Vocabulary (Monogram Language Model - non context based), are relative to the underlying phonetic pronunciations. When you train an individual word, the underlying phonetic representation of the pronunciation is adjusted accordingly, but the probability coeficient of any word, other than custom words that do not exist in the background vocabulary list, remains the same. Probability coeficients are modifiable only for the bigram and trigram language models (context)and custom words. The reason for this is too complex for discussion in this post. Suffice it to say that there is a critically important reason for this.

Fourth, when correcting phrases, DNS applies the appropriate Language Model based on the number of words:

1. One word - Vocabulary (Monogram LM)
2. Four words or less (Bigram LM - context analysis)
3. Five words or more (Trigram LM - context analysis)

It is sufficient to note here that invoking the Trigram LM will result in greater accuracy, as well as the speed of display, than the Bigram or Monogram LM's. This is why training and correction of phrases of 5 or more words is more effective at optimizing accuracy. Explaining why beyond this is again complex beyond the scope of this post.

Fifth, if you have the Automatic playback on correction feature turned on, and you do not hear automatic playback of your original dictation of the word or phrase being corrected, this means that there is no underlying related to speech in the speech buffer. This can be because you made a correction manually to the text containing this word or phrase, or, for some other reason, the text buffer was flushed. As I stated in a previous post, this is a tricky condition. Sent you have no playback of your original spoken dictation of this word or phrase, you have to decide whether or not you said to correctly in order to decide whether or not you should train it. Under this condition, if you elect to train your correction, or the correct selection option, only that option will show up in the subsequent training dialog. If there was playback, one way of determining this is that if you select a correction option, or type or dictate the correct word or phrase in the text box at the top of the Spell dialog, and then select train, both the correction and the misrecognition will show in the training dialog box. The recommended procedure that I generally use is to assume that what you said originally was correct (your spoken form), and simply select or type the correct option and click OK without training. This is usually the most appropriate, effective, and optimal choice.

Lastly, never, and I repeat never, use the correction function to change a word or phrase that is correct (i.e., not a misrecognition). Select it and redictate what you wanted to say (Select-and-Say). Doing so can result in entering contradictory and conflicting links between the Acoustic Model and Language Models. The result can be total misrecognition of the words or phrases involved, and is likely one reason why accuracy degrades significantly for some users.

#1 is the word or phrase that was displayed during the original dictation. If it is correct, then you are not correcting a misrecognition, you are correcting a word or phrase that was initially correctly recognized. If you look carefully at the options in the Spell dialog or Correction window, you will notice that selection #1 has no words in bold. This means that these were the words that were displayed when you originally dictated that word or phrase. If they are incorrect, than the correct word or phrase should be between #'s 2 and 9. They selection options will have at least one word bold, which indicates the correction that will be made to the original word or phrase. In short, if #1 in the selections is the correct word or phrase, that it was never misrecognized to begin with. However, this is again another reason why the playback option is important in the proper use of the correction function.

There are numerous other training and optimization techniques that augment these and improve the accuracy of DNS, but these are better left to a subsequent post.

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