When you select to Edit a transcript in the transcript editor, you may notice a Word Confidence slider above the transcript text pane as shown in the below figure. The confidence slider appears for any transcript generated by the Amazon ASR automatic transcription service (unless it has been previously edited as described later on this page).
What is a confidence score
When Amazon performs a transcription, each word is given a "confidence score" based on how certain the machine transcription was in selecting the correct word. Lower confidence often indicates speech that was garbled, difficult to understand, lower in volume or other reason it may not have been interpreted properly by the automated program. The scores are presented as a percentage and as such range from 0 to 100.
Words are underlined in the transcription panel based on the confidence value configured for the slider. The value shown by default is an average confidence score across the entire transcript. Underlined words are those whose scores fall below the confidence score.
Changing the confidence score applied
Click and drag the Confidence slider (or type a different value in the box located to the right of the slider) to CHANGE the confidence level used to underline words in the transcript. For example, if the speaker had a thick foreign accent, you may want to lower the confidence percentage to reduce the number of underlined entries. This lets you target the words most likely to have been incorrectly transcribed.
Lowering the confidence slider a bit, particularly for foreign speakers, is a good way to find those terms that are repeatedly mis-transcribed, and use the Search/Replace functionality to fix them. Then raise the slider to find remaining words and review them.
NOTE: Word Confidence is NOT the same as accuracy! While we wish machines could provide us with 100% fully accurate transcripts, at this time the best they can do is gauge how confident the algorithm is with its output. Word Confidence is a good tool to first review words that may be inaccurate, but it should not be the only indicator and is not an alternative to reading through the entire transcript and confirming its accuracy against the media's audio.
Do all transcripts have confidence scores
All transcripts returned from the Amazon ASR transcription service will have a confidence score applied to each word. These are contained in the NOTE CONF entries of the transcript file (see Defining sections of a WebVTT file in Editing Transcriptions Outside of Echo360 for more information). If you upload an automated transcript that you may have edited offline previously, and the file still contains the NOTE CONF fields, the confidence ratings will be used for the Transcript Editor interface.
If a transcript file is uploaded that does not contain the NOTE CONF fields, the Word Confidence slider is grayed out and the value shown in the box is 100%. Essentially, the editor assumes that if there is not a confidence score, that the transcript was generated by a human and is therefore correct.
When a user makes their edits, the confidence scores are CLEARED from the entire cue (single line of words) where ANY of the words have been edited. Meaning that if you edit a single word in a cue where multiple words are underlined for being below the confidence score slider threshold, when you move to a different cue, the underlines in the entire cue are removed, because the confidence scores are cleared.
As indicated above, no confidence score implies 100% because human intervention is assumed. In this case, it is assumed that the entire cue was reviewed along with the change and as such, the cue is 100% correct. While the interface will show that the scoring has been removed for each edited cue, you still have to save the edits to a new version, at which point the scores are removed from the transcription.