I am a new user of Dragon NaturallySpeaking, and I would appreciate some expert guidance about when and how often I should perform training. I am using the UniVoice profile, and have added as custom words a number of words and phrases that I use at work.

When I used the Accuracy Center to add those words and phrases, the default option was also to train them, which I did. Was following the default a good idea, or are there other factors that I should consider?

Also, is it appropriate to use the KnowBrainer “Quick Train" command to train utterances that I have corrected? Is this something that is desirable to do frequently, rarely, or somewhere in between?

Finally, is it an acceptable practice to train a fully formatted phone number by saying only the numerical digits? For example, is it okay to train (410) 857-8777 as a single utterance of '4108578777'?

I have seen a lot of good discussion here about proper correction technique, and I am hoping that folks will also share their expertise about that training practices as I work to optimize the accuracy of the UniVoice profile for my speaking voice and for my writing situations.

This is a wonderful forum, and I am grateful to have discovered it can to have the opportunity to interact with and to learn from  so many experts.

Andy.

-------------------------

DNS 12 Professional with KB 2012, VoiceComputer 12, and Samson Airline 77.  

Running on Maingear Shift Desktop: Intel Core i7 3930K (Sandy Bridge-E) overclocked to 4.2G5Hz, with 256KB X 6 L2 and 12MB L3 Cache; ASUS Rampage IV Extreme Motherboard, 64GB Corsair Dominator DDR3 Memory;  2 X Nvidia Geoforce GTX-580 display adapters; 8 X Samsung 830 SSDs arrayed as 512GB R0 (OS) and 713GB R10 (DATA).

The 1st rule of thumb is the path of least resistance. In other words avoid any real work whenever possible. Think lazy.

Unless you are adding a word that is unique to DNS (not in the current or back up dictionary) we don't typically recommend training words. We never recommend training phrases because it's very rarely necessary.

The KnowBrainer Quick Train command is designed to train a word in context (with other words). We believe that it is almost a complete waste of time to train a single word from within Spell That correction. If you wish to train individual words, we only recommend doing so from within the Vocabulary Editor but we don't recommend a lot of training and we never recommend running extended training sessions unless you are having a serious accuracy problem. In this particular case, we only recommend one additional training session. More than 1 training session can actually damage your user profile. 

-------------------------

Click KB 2012 REV D to Download a 30 Day Evaluation of KnowBrainer 2012 

 

 

 

Okay, so if I understand correctly, the best practice is only to train entirely new words that you enter into the Vocabulary Editor, but otherwise rely on the normal correction tools without additional training. I imagine that, if there are specific words that do not correct well, that would be appropriate to use the Quick Train command on those individual words-but only sparingly. So, the basic message is don't be profligate with training.

 

Andy.

-------------------------

DNS 12 Professional with KB 2012, VoiceComputer 12, and Samson Airline 77.  

Running on Maingear Shift Desktop: Intel Core i7 3930K (Sandy Bridge-E) overclocked to 4.2G5Hz, with 256KB X 6 L2 and 12MB L3 Cache; ASUS Rampage IV Extreme Motherboard, 64GB Corsair Dominator DDR3 Memory;  2 X Nvidia Geoforce GTX-580 display adapters; 8 X Samsung 830 SSDs arrayed as 512GB R0 (OS) and 713GB R10 (DATA).

Andy,

I've added about 18.000 custom words to the vocabulary over the years, and I don't think that I've trained more than ten of these words.

So basically -- don't train!

Unless it's a word that NaturallySpeaking keeps getting wrong on and on. Even then, correction without training, or adding the word in context as a phrase, usually will help more than training.

Cheers David.P

-------------------------

Core i5 @ Windows 7 64bit
Sennheiser MKH HF Capacitor Desktop Gun Mic @ M-AudioTransit USB
Fingerworks iGesture Multitouch Pad & 3DConnexion SpaceNavigator + Visual & Acoustic Feedback + Automatic Microphone Control
NaturallySpeaking installed in SATA RAM SSD Solid State RAM Hard Drive

 

 

Download NaturallySpeakometer and objectively measure the perfomance of your NaturallySpeaking installation, compare the impact of different hardware (microphones, processors, hard drives etc.) on your speech recognition system speed -- also compatible to Vista, Windows 7 and Windows Speech Recognition

[edits necessary for this forum not having a "preview your post" function]

When you say "correction without training", do you mean selecting a phrase that has an incorrect word (to bring up the correction box), and either choosing the correct selection from the list and if the correct word is not there typing it in under "Spell That"?

 if I understand correctly, that should be virtually the only way training takes place.

 Also, when you say you have added 18,000 words, is this done through the above process? Are new words you type in automatically added to the dictionary?

 

Thanks,

 

Howard

That's exactly what I do.



As said, I almost never train anything, because usually correction is all it needs to improve your Language Model as well as your Acoustic Model. On many occasions training even will be entirely futile -- that is, every time your pronunciation doesn't quite match with NaturallySpeaking's (internal) automatic word phonetization, NaturallySpeaking will simply ignore your training.



Yes and yes -- if you enable "automatically add words to vocabulary" in the DNS options which you definitely SHOULD do starting with DNS SP1 since the "automatically add crap to vocabulary"-bug after (almost a decade) finally has been fixed with SP1.

Cheers David.P


-------------------------

Core i5 @ Windows 7 64bit
Sennheiser MKH HF Capacitor Desktop Gun Mic @ M-AudioTransit USB
Fingerworks iGesture Multitouch Pad & 3DConnexion SpaceNavigator + Visual & Acoustic Feedback + Automatic Microphone Control
NaturallySpeaking installed in SATA RAM SSD Solid State RAM Hard Drive

 

 

Download NaturallySpeakometer and objectively measure the perfomance of your NaturallySpeaking installation, compare the impact of different hardware (microphones, processors, hard drives etc.) on your speech recognition system speed -- also compatible to Vista, Windows 7 and Windows Speech Recognition

[edits necessary for this forum not having a "preview your post" function]

Thanks.

 

By the way, I turned off the auto add because I thought that allowed the same problem as in the past. Thanks for the tip, I'll turn it back on.

 Howard