Is it possible to bias the decoder toward a small set of tokens?

[tilsen]

For example, if I wanted to transcribe audio in which there is a small set of words, is it possible to bias the decoder model to prefer those words over similar ones, so that transcriptions become more likely to output those words?

[glangford]

Try the --initial_prompt option, and fyi below.

• Is it possible to do model adaptation? #66
• Can I use a prompt with this whisper version? #1159
• https://github.com/BriansIDP/WhisperBiasing