feat: bind whisper.cpp Parakeet API

[jatinkrmalik]

Summary

Adds Python bindings for whisper.cpp's Parakeet API (parakeet.h), which is a separate C surface from Whisper.

• Upstream: ggml-org/whisper.cpp#3735.
• Requested in #171 for Vocalinux #527.

What this does

1. Bumps the whisper.cpp submodule to 080bbbe85230f624f0b52127f1ae1218247989f9 (master after Parakeet landed; previously v1.8.4 / 9386f239).
2. Links _pywhispercpp against the parakeet library as well as whisper.
3. Exposes a minimal C API through pybind11 (src/parakeet_bindings.cpp):
   • parakeet_context_default_params / parakeet_init_from_file_with_params / parakeet_free
   • parakeet_full_default_params / parakeet_full (float32 PCM)
   • segment count, text, t0/t1
   • parakeet_version, PARAKEET_SAMPLE_RATE
4. Adds ParakeetModel (pywhispercpp/parakeet_model.py): load a local GGUF, transcribe() path or float32 mono 16 kHz PCM, returns the same Segment objects as Whisper Model. Audio path loading reuses Model._load_audio. No automatic model download; models from ggml-org/parakeet-GGUF.

Day-one scope matches the issue: load model, run audio, get text. Not every parakeet-cli flag.

Out of scope (follow-ups)

• Streaming / server path
• Full token-level callback parity with CLI
• Automatic download helper for Parakeet GGUFs

Test plan

• Build extension with updated submodule (local cmake/ninja on Linux)
• Unit tests that do not need a large model download (tests/test_parakeet.py: defaults, version, free/None, import, missing path)
• Manual smoke with a Parakeet GGUF from ggml-org/parakeet-GGUF
• Confirm existing Whisper Model tests still pass in CI (needs model files / full package install)

Closes #171

[absadiki]

Thanks so much @jatinkrmalik for the early draft! I really appreciate it.

This looks like the direction I had in mind. Having a separate ParakeetModel while keeping a similar high-level Python interface makes sense, since Parakeet exposes its own native API (parakeet.h)

I gave the PR a quick try locally, but it didn't work at first. I noticed that Parakeet support wasn't actually introduced until whisper.cpp v1.9.0, so we'll need to bump the submodule to at least that version.

While looking into it, I realized the project could benefit from some refactoring to better support additional model architectures in the future. So I went ahead and implemented those changes on top of your work.

Whenever you get a chance, could you try #173 on your end and let me know if everything works as expected? If it does, we can use that as the basis for getting Parakeet support merged.

One small note: the old Model class is now deprecated, though it will continue to work for backward compatibility. Going forward, I'd recommend using the new WhisperModel class. I also switched the parameter handling to dataclasses, which provides much better IDE autocomplete, type checking, and overall usability.

Feel free to take a look through the code, I think the new structure should be fairly straightforward, and it should make adding support for future model architectures much cleaner.