Transcribing uncommon words with whisper

[OthmanePreure]

I'm transcribing railway related audio files, however the model gets most of the Time the name of the train stations wrong
I tried fine tuning whisper with audio files (2 hours) but the results wasn't good enough. then i tried with prompt but the size of the context is limited to 244 token which in note sufficient regarding the important number of stations.

Any tips to solve this issue (maybe using some methods with post processing)

[glangford]

Post-processing with soundex or metaphone or some other form of fuzzy matching? eg. checking Whisper generated spellings against a dictionary of properly spelled station names

Phonetics based Fuzzy string matching algorithms - A dive in Soundex & Metaphone

https://medium.com/data-science-in-your-pocket/phonetics-based-fuzzy-string-matching-algorithms-8399aea04718

[rahulbansal16]

@OthmanePreure Did you try the phonetics-based matching?