When you are leaning on Limecraft for transcription purposes using speech-to-text technology or AI transcription, you can use the confidence score to rapidly spot areas that may contain errors. This article explains how to understand and use the confidence score.
TABLE OF CONTENTS
Instructions on how to use the confidence score
Before you begin, make sure you have a registered account on Limecraft and that it is properly provisioned with transcription credits.
What is the Confidence Score of machine transcription?
As a result of the machine transcription or AI transcription process, the Automatic Speech Recognition (ASR) engine attributes a confidence score to each individual word. The confidence score is a percentage between 0% and 100% that represents the probability of correct recognition. The lower the confidence score, the lower the probability of correct recognition.
The reason for using confidence scores has to do with the strategy by which the ASR engine tries to understand the spoken word. In particular, it is trying to detect sentences rather than individual words. In case there is not a perfect match due to a strange language construction, the use of dialects, or the use of words in other languages, this will be indicated by a lower confidence score for such words.
How can you visualise the confidence score?
Assuming you successfully uploaded and transcribed a clip of audiovisual material, you can toggle between the standard view, highlighted words that have bad confidence scores, and highlighted words with good confidence scores.
? If you are looking for potential errors, you should highlight bad confidence scores. The lower the confidence, the more dark and red.
How can you use the confidence score to speed up the post-editing process?
While the clip above overall has little or no errors, as no more than one word with low confidence is highlighted, other clips or certain areas in other clips can be more problematic. In your capacity as a transcription operator ("expert in the loop"), upon successful completion of an automated transcription process, you want to be able to quickly scan the transcript so as to identify those areas.
In the example below, the primary language of the clips was English and the clip was correctly transcribed using English ASR. However, the English speakers are interrupted by a French speaker ("Male 3"). Obviously, the result of English ASR is complete nonsense, which is visualised as an area where the majority of the words have low confidence scores.