Transcription from a Device

Chuck Rogers theChuck at mac.com
Thu Dec 26 17:25:31 PST 2002


All:

Here is an excerpt from my upcoming book, "The Macintosh Speech 
Recognition Handbook" that covers this topic. This excerpt assumes you 
have a microphone with a connector that fits into your recorder. (There 
are more detailed instructions on selecting microphones, and USB 
adapters, etc. elsewhere in the book.)

Create a “Digital Recorder Profile”
Since the input from a recorder sounds different to the software than a 
live person, you should actually record the training stories on the 
recorder and create a trained "digital recorder" profile in order to 
get the best accuracy. This is going to require some work on your part. 
It is a laborious procedure in which only those with an intense desire 
for transcription should partake.

Here’s how to do it:

1. Attach the microphone you will be using for dictation to your 
recording device. You can use the recorder’s built-in microphone, but 
you may not get very good recognition.

2. Pick up a Y-adapter from Radio Shack and route one cable from the 
mic to the recorder and one to your computer.  You can set your 
computer to use its built-in microphone, but your dictation software 
may have trouble with accuracy while reading the stories. If you have 
two of the same headsets you can use one for the recorder and one for 
your computer. Wear one around your head and the other around your 
neck, with the microphone boom on the opposite side.

3. Launch your dictation software and create a new profile. Name it 
“Trash Me” or something similar (it doesn't matter, since you won’t be 
saving this profile).

4. Run the “Set Up My Microphone” procedure (iListen) or “Audio Setup 
Assistant” (ViaVoice). Make sure you record the text presented into the 
recorder.

5. Run the “Learn My Voice” procedure (iListen) or “Analyze My Voice 
Assistant” (ViaVoice). Make sure are recording while you are reading 
the text.

6. Continue reading the training stories, recording them into the 
recording device. Stop the recorder between each story to create a 
separate session for that story. There are 12 stories in iListen and 6 
in ViaVoice. (NOTE: This may be too many stories for the available 
recording time on some recorders. In that case, record the microphone 
setup procedure, ONE training story, and the sample text described in 
the next step.)

7. Read some very simple and easy sample text of about 300 words into 
the recorder. It is important you read the words, since reading will 
sound a lot like the training stories to the program. Choose text that 
has very few technical terms or proper names that the speech 
recognition software is unlikely to know.

8. Hook up the recorder to the sound input of your Macintosh. You may 
need Griffin’s iMic adapter if you have a USB only Macintosh.

9. Launch your speech recognition software (if it is not already 
running) and create a new profile. You can call it something like 
“Transcription.”

10. When the Microphone Setup procedure appears, play the session with 
the appropriate text. You may have to rewind and continue playing the 
sample if the program has trouble with the device for some reason.

11. Continue to the training stories portion of the software, playing 
the appropriate file from the recorder as the text appears on the 
screen. This will require some dexterity on your part since you will 
need to manipulate both the program on screen, and the recorder. It may 
be necessary to rewind and play certain words over again, and this may 
require some practice to get it right. Feed only one training story to 
the software and allow it to analyze what it heard.

12. Now activate the dictation part of the software and feed it the 
Sample Text session from the recorder (or from the file on the hard 
disk, if you are using iListen's Transcription feature.). In ViaVoice, 
have the program type the text into the SpeakPad. With iListen, have it 
typed into SimpleText (Mac OS 9) or TextEdit (Mac OS X).

13. Determine the accuracy by counting the number of misrecognized 
words or phrases. Don’t count proper names or technical terms. Also, 
count misrecognized phrases as one error (see Chapter 5, "Training 
Camp" for more information on determining accuracy).

14. If accuracy is below 92%, feed another training story to the 
program (you will have to reactivate its training feature), analyze, 
and then feed it the sample text again. Continue doing this until 
accuracy is at least 92%, at which point you can use the respective 
program’s Correction features to further improve your transcription 
profile.

The book is in the process of being edited and illustrated now, and is 
slated to be published sometime this spring.

Best Regards,

Chuck Rogers


More information about the MacVoice mailing list