WER and use cases for strict and subtitle versions
Is the release of WER for strict and subtitle versions on your roadmap? I am interested in seeing the comparison of all 3 versions. Secondly, what specific usecases do strict and verbose suit better?
Hi! Please find the updated tables with WER and BLEU evaluated for each of the three model versions (standard, subtitle and strict) at the bottom of each model card https://huggingface.co/KBLab/kb-whisper-large#evaluation. The subtitle version is very good at shortening the output, the strict one is more verbose, and standard is somewhere in between. The strict version suits use cases where it is important that all words are represented in the transcripts, for example in protocols or interview transcripts. But that also varies depending on the use case, so I suggest you try the different ones and find the verbosity you prefer.