STM is primarily for command and control, those 30 or so things that the user does all the time: open, save, menu bar operation, copy, paste, go to a certain web page, launch a program. You can even control the cursor by voice.
You can have an AAC device control the computer, by voice output! It's quite cool. Since STM doesn't really care if the user sounds like how he/she is 'supposed' to sound, it can respond to the sound output of the user's AAC device! Just have the device near the computer microphone and it's almost perfect!
And we don't 'average' the 3 trainings; we record all three. So the user could actually train it 3 different ways so as to maximize the probability of recognition during use. However, I recommend trying to train consistently and recall that way for recognition.
People ask me about 'dictation' and I always recommend NS, when possible.
However, a workaround to use STM for dictation is to train the alpha, bravo, charlie, etc. alphabet and use in conjunction with Co:Writer for word prediction.
You can even set up Co:Writer in scan mode so it auditorially scans through choices, for blind users.
For further information visit rjcooper.com/speak-to-me/