Maximum number of 'initial_prompt' characters/tokens in transcribe()?

[rmohl]

Hello!

I was wondering what might the maximum number of characters/tokens that can be given in the initial_prompt when using transcribe() be before they stop affecting the transcription?

Just a note: I forked the code and altered it to fix the issue described in this comment. Essentially the initial prompts now affect the entire transcription, rather than just the first part of it.

In the comment refered to above, the poster says:

Even if I pass in initial_prompt, it gets concatenated with the previous window transcriptions which get truncated to the model's model.dims.n_text_ctx // 2. For the tiny model, that's the last 223 tokens only

Would this value maybe be the max number of tokens I could use for initial_prompt? Or is it more complex than that?

Thanks!

[SeeknnDestroy]

hi @rmohl could you find your answer? would you mind if you share your thoughts on your experiments otherwise?

[rmohl]

I ended up just assuming that the max number of tokens you can use for initial_prompt is 223. According to this OpenAI website approximately every 4 characters of text is equivalent to a single token. So the max prompt length (for my altered version of Whisper at least) is about 4 * 223 = 892 characters.

[SeeknnDestroy]

thanks 🙏

[JenuelDev]

by the way what does the prompt do?