Transcription Feature

natrona111 wrote: I am most interested in the "Transcription" service that WSRTookit can provide.  I have tested it out on a few wav files that I have of recorded conversations from a call center.  A few things I have noted - 1) the program doesn't seem to be transcribing the entire conversation (it will stop transcribing about halfway through a 5 minute conversation) and 2) the accuracy rate is about 10%.Any suggestions on improving this feature?

You do not seem to have an understanding of WSR let alone speech recognition in general. WSR, Dragon NaturallySpeaking, IBM are examples of large vocabulary speech recognition software. These software are what are called speaker dependent. They require several things. Among these:
1. The software must be trained to the voice of the single individual who will be using it.
2. You cannot speak conversationally. Every word must be enunciated clearly. This is because the software first listens for the sounds of each word. If you slur words together the software will work poorly even under the best conditions even with the best microphone.
3. The next step in the process is each word is compared to the words before each word and the words after for context clues. This is why you must dictate punctuation marks like periods and commas to help the context.
4. It helps if you dictate phrases as this facilitates the software finding the context of words.
5. Finally, when you make corrections of your mistakes, the system learns to not make that mistake again (usually).

WSR and other speech recognition software have never been able to transcribe lectures with any degree of accuracy. This is because lecture is voices have not been trained to the system. They generally do not enunciate as clearly as is necessary for best accuracy. They do not use punctuation and the text is not corrected in order to improve accuracy.

The best way to transcribe lectures is to get something like a foot pedal to play a little bit of a lecture into earphones, use another pedal to pause the recording, use WSR or Dragon NaturallySpeaking to dictate into your word Processing Software and then play a little more of the lecture and so on.

Marty Markoe, eMicrophones, Inc.