Test of Audio Transcription via Dragon Naturally Speaking by Diane Chapman (SPHS)
Background
The need for looking at ways of providing accessibility to those people with disabilities who take our online courses has arisen. Much of the content in our online courses consists of PowerPoint presentations synched with streaming audio (aka tutorials). As a result, we needed to test the ability of a voice recognition program to transcribe previously recorded audio. Following are the results of one such test.

Procedure
I had recorded a short (18 slide) PowerPoint presentation earlier this semester and had it made into a tutorial. To perform this test I needed a copy of the audio of this presentation. Since the audio is cut up during the editing process to synch the slides, the only audio of the full presentation I found was a Real media version of the unedited recording. This is the recording that was used to perform this test.
I also used Dragon naturally speaking as the voice recognition program The specs are Naturally Speaking Preferred, version 5.

• Preparation of the Audio
  Dragon Naturally Speaking (DNS) has a function for transcribing audio, but the audio must be either a DNS or .wav file. Since the only file I had was in real media (RM), I had play the (RM) file and rerecord it as a .wav file. The result was a .wav file (24,497 kb).

  Training the Voice Recognition Program
  In order to transcribe an audio file, DNS must be trained to recognize the particulars of a person's voice. DNS provides a series of specific readings that must be read into the program via a microphone. The more reading that are done, the better DNS will be at distinguishing words correctly.

  The process for this test was to train DNS four time and to assess the correctness of the transcription after each time. Each training is supposed to improve the accuracy of DNS. Training DNS for my voice took time. The time for each training was as follows (also noted are the specific texts that were read at each training):
  Training 1: 10 minutes (Talking to Your Computer)
  Training 2: 15 minutes (Alice's Adventures in Wonderland)
  Training 3: 5 minutes (Talking to Your Computer)
  Training 4: 7 minutes (What's Plain English)

  Transcription of the Audio
  Because the audio recording was unedited, there are several places where slides were redone and extra audio was captured. All of the test results have been edited (using correction tracking in MS Word) so that you can see exactly where the problems are.

  I set myself up as a user and trained my voice for the first time (taking about 10 minutes). The results of this test are saved in DDC_test1_edited.doc.
  I trained my voice for the second time (taking about 15 minutes). The results of this test are saved in DDC_test2.doc.

  I trained my voice for the third time (taking about 5 minutes). The results of this test are saved in DDC_test3.doc.

  I trained my voice for the fourth time (taking about 7minutes). The results of this test are saved in DDC_test4.doc.

  Test Results
  The edited transcripts from all four tests were reviewed and the number of errors counted. The results are as follows:
  Test 1: 75 errors (See DDC_test1_edited.doc)
  Test 2: 87 errors (See DDC_test2_edited.doc)
  Test 3: 75 errors (See DDC_test3_edited.doc)
  Test 4: 92 errors (See DDC_test4_edited.doc)
  

Special Problems & Issues:
By performing this test, I have found out several pieces of information that are important.

1. We can transcribe pre-recorded audio using a voice recognition program. But, although the process is possible, it is problematic.
2. While voice recognition has come a long way, its accuracy is still a problem.
   The subsequent trainings of my voice seemed to have little or no impact upon the accuracy of each subsequent transcript. Still, 75 errors in an 18-slide presentation is excessive. The subject matter expert will be the one who will have to review and make corrections to the transcript. This is not practical or realistic to think that a faculty member will do this when we ware talking about hour-long presentations with 75 or more slides.
3. The audio must be re-recorded into a .wav file. While this is not a problem, it does take extra time and I'm sure that the level of audio quality decreases.
4. The transcript is produced void of any punctuation. This means that even if the transcript is error-free, someone must manually go through each text presentation and capitalize letters and add punctuation. This is very time consuming and should optimally be done by a content expert.

As a result of these limitations, I do not think that the transcription function in Dragon Naturally Speaking will fit with our needs for audio transcripts. It is a very labor-intensive process and requires skilled manpower already in short supply. However, it may be useful for those giving presentations to use it to prepare their own scripts prior to recording in the audio booth. They could give their lecture at their desktop, edit it and give us a copy when they give us their PowerPoint presentation.