Context
whisper.cpp added NVIDIA Parakeet support in June (ggml-org/whisper.cpp#3735, around v1.8.7 / v1.9.x). It is not another Whisper ggml file you pass to whisper_init_from_file. Upstream exposes a separate C API (parakeet.h) plus parakeet-cli.
Converted models live under ggml-org/parakeet-GGUF (for example ggml-parakeet-tdt-0.6b-v3-q8_0.bin).
Why I'm asking
I maintain Vocalinux, a Linux dictation app that uses pywhispercpp as the default engine. We got a feature request to offer Parakeet in Settings (vocalinux#527).
I tried to see if this was just a model-list change on our side. It isn't. pywhispercpp.Model only wraps the Whisper path, and the packaged whisper.cpp is still pre-Parakeet (1.4.1 ships ~1.8.2; 1.5.0 also predates the merge and still binds whisper_* only).
Request
Would you be open to:
- Bumping the whisper.cpp submodule far enough to include Parakeet
- Binding enough of
parakeet.h for the common case: init/free, parakeet_full (or equivalent), segment text, threads / GPU knobs
- Optionally a small high-level helper similar to
Model.transcribe(), if that fits the project
I do not need every parakeet-cli flag on day one. Load model, run PCM float audio, get text back would already unblock us.
Notes
- Server support is not in whisper.cpp yet, so this is about the in-process API only.
- Happy to test a branch or pre-release wheel from our app and report back.
Thanks for maintaining the bindings. They already do a lot of heavy lifting for us.
Context
whisper.cpp added NVIDIA Parakeet support in June (ggml-org/whisper.cpp#3735, around v1.8.7 / v1.9.x). It is not another Whisper ggml file you pass to
whisper_init_from_file. Upstream exposes a separate C API (parakeet.h) plusparakeet-cli.Converted models live under ggml-org/parakeet-GGUF (for example
ggml-parakeet-tdt-0.6b-v3-q8_0.bin).Why I'm asking
I maintain Vocalinux, a Linux dictation app that uses pywhispercpp as the default engine. We got a feature request to offer Parakeet in Settings (vocalinux#527).
I tried to see if this was just a model-list change on our side. It isn't.
pywhispercpp.Modelonly wraps the Whisper path, and the packaged whisper.cpp is still pre-Parakeet (1.4.1 ships ~1.8.2; 1.5.0 also predates the merge and still bindswhisper_*only).Request
Would you be open to:
parakeet.hfor the common case: init/free,parakeet_full(or equivalent), segment text, threads / GPU knobsModel.transcribe(), if that fits the projectI do not need every parakeet-cli flag on day one. Load model, run PCM float audio, get text back would already unblock us.
Notes
Thanks for maintaining the bindings. They already do a lot of heavy lifting for us.