Whispercpp v1.8.4 api

[scottmonster]

So.... I think I messed up in during that last PR for typing. When I added **params to typing I pulled just to make sure there were no conflicts from the other PRs. However I think I pulled origin (my fork) instead of upstream BEFORE my origin/main had the updated v1.8.4 whisper.cpp. So when you merged that typing branch it also reverted to the original whisper.cpp version. No big deal because this PR includes it. I just wanted to give you a heads up.

CHANGES:

• fix prompt_tokens binding: Binding was just a pointer but constants.py indicated tuple.
• change Model.__call_new_segment_callback from a static method to an instance method so that self._new_segment_callback resolves the instance callback not a global callback (ie you can have multiple model instances with different new_segment_callbacks)
• change Model.transcribe so that new_segment_callback is not persistent. when new_segment_callback == None, it will always make sure the callback is cleared
• add abort_callback to Model.transcribe, it has the same behavior
• update whisper.cpp to 9386f239401074690479731c1e41683fbbeac557 whisper.cpp (v1.8.4) ... again
• main.cpp bindings better integrate with whisper v1.8.4 (specifically context params)
• bindings for all callbacks now have same shape... mostly
• Add a compatibility alias that maps suppress_non_speech_tokens to suppress_nst internally.
• remove all WHISPER_DEPRECATED functions as defined by whsiper.h
• update README.md via AI... I looked over it didn't see any issues.

OTHER ISSUES/TODO:

• misc. type issues (model.system_info() -> None: should probably return str)
• bindings for callbacks could probably use a template
• move all of the typing to a single place so it doesn't have to be maintained on multiple places
• possibly add default new_segment_callback and abort callback to the model class

If you care to see some of the documents and scripts I used you can look at my working branch. Specifically the coverage and tests directories. A lot of that is AI generated and didnt feel like it substantially improved anything so I left it out. Although you might be interested in having tests/compat and tests/run-compat.py

[absadiki]

No worries at all, thanks @scottmonster for the heads up and for the PR.

This looks like a really solid cleanup/update overall.

One thing I noticed is that the default language was changed to "en". I think it should remain empty str so whisper.cpp can continue auto-detecting the language by default. wht do you think ?

Also I noticed the AI README changes seem to introduce quite a bit of formatting noise (extra empty lines / spacing changes) without much actual content improvement. I think it would probably be better to keep the README diff smaller and focused on meaningful changes only.

Also, feel free to include the additional test files/scripts you mentioned if you think they improve coverage or compatibility testing. More tests are always welcome when they provide useful coverage.

[absadiki]

I think this should remain "" for auto-detecting the language by default.

[scottmonster]

ah yes, you're right. Originally I was trying to match the default whisper-cli args which defaults to en. Once I realized I was going about it the wrong way I had to change a bunch of stuff. I'll change it back to an empty string and provide a cleaner README shortly

[absadiki]

Ahh, gotcha .. no worries.

[scottmonster]

So I literally just copy and pasted your README back in and changed line 295 to

ctx = pwcpp.whisper_init_from_file_with_params(
  'path/to/ggml/model',
  pwcpp.whisper_context_default_params(),
)

[absadiki]

I think everything looks good so far .. we can merge this for now.
Thanks again @scottmonster for the work on this PR.