[j-say list] Re: Getting accurate speech-recognition

  • From: "Gary Wunder" <GWunder@xxxxxxxxxxxxx>
  • To: <j-say@xxxxxxxxxxxxx>
  • Date: Thu, 31 May 2012 17:26:33 -0500

Talk more about item 4 please. How does one easily import the things
customized such as commands and vocabulary?

 

-----Original Message-----
From: j-say-bounce@xxxxxxxxxxxxx [mailto:j-say-bounce@xxxxxxxxxxxxx] On
Behalf Of Pranav Lal
Sent: Wednesday, May 30, 2012 7:46 PM
To: j-say@xxxxxxxxxxxxx
Subject: [j-say list] Getting accurate speech-recognition

Hi all,

This message is the result of a post by Harry Brown where he mentioned
inaccuracy using speech-recognition. I have elected to create a separate
thread for this since there are many factors that could lead to
speech-recognition inaccuracy. I am going to try and address the issues I
have encountered personally and those I have read about. I am not picking on
anyone here.
 1. The language model
It is important to understand how Dragon works. Dragon Naturally Speaking
uses Markov chaining for speech-recognition. As it says in the basic
training, Dragon looks at the context in which words appear. So, if cat and
hat usually appear together and you dictate cat and bat Dragon will assume
that you meant cat and hat. The way you get around this is using the
correction system. So if you dictate cat and bat but Dragon has written cat
and hat, you do the following.
A. You position the insertion pointer anywhere on the utterance that
contains the words cat and hat. The way you do this is to move the insertion
pointer word by word and place it on any of the 3 words cat and or hat.
B. You then hit f10 or say "correct that".
C. You check the suggestions list or use the correction box commands to
change hat to bat.
D. You then exit that box by saying "confirm action" or "confirm that" or by
hitting enter.

This teaches Dragon that in your writing style, cat and bat occur together.
Always play back your dictation to ensure that there are no errors in your
utterance.

2. The vocabulary
Dragon comes with a default vocabulary that is quite large. I believe it is
60000 words but I cannot be sure of this. If you use words that are not in
the vocabulary, you need to add tem. Otherwise, Dragon will not understand
them. If you plan to use the words only once, you can get away by spelling
them out. The way you check if a word is in the vocabulary is as follows.
A. You launch the vocabulary editor.
B. You search for the word by typing it in.
C. If you find the entry for that word, you are good to go. If not, you need
to add the word. I will cover the vocabulary editor in a podcast that I plan
to do  on this subject. I hope to get to this by the middle of June.
D. You can train the pronunciation of this word from the vocabulary editor
too.

When it comes to adding words, you can use the "my words" feature or add
them directly into the vocabulary editor. Alternatively, you can make a list
and use the accuracy center to add them. Please do not give documents to
Dragon for analysis since it is possible to confuse dragon. Give it lists of
words in plain text file so that you avoid formatting which could again
confuse dragon.

3. How you speak
You need to speak like a newscaster so no getting excited.

4. The hardware
This is where many installations fall  over. You need a sound source that is
free from any kind of noise. This includes computer fan noise etc. The
reason everyone suggests that you go with external sound cards is because
many computers have poorly shielded components so noise or interference
could bleed through into your recording.

Diagnosing inaccuracy causes
1. Check how you are speaking. Play back your dictation and compare it to
what you hear on the news or on the material that T&T have supplied. There
is a recording of Brian Hartgen speaking the training text for using a
digital recorder. Use that as a guide. Do not try to sound like him. Please
retain your original voice and accent but mark his intonation or lack of it.
<chuckle
2. Check for interference. When you hear your dictation, look for background
hums or any other noise. Is your speech-synthesizer being picked up by your
microphone?
3. Check the vocabulary. Are your words in Dragon's vocabulary editor?
4. Could your user profile be corrupted? Many of us get attached to our user
profiles. I personally just archive my macros and special words and blow it
away. Create a fresh user profile and import your words and commands. See if
that helps.

I hope this helps someone. Please read the documentation that comes with
Dragon to get a better understanding of all of this. T&T have done a good
job of simplifying things but when things do not work we do need to go
beyond what has been written and explore.

Pranav



Other related posts: