Support for Specific German Parkakeet Model

[MyButtermilk]

Currently the normal Parkaeet Model is supported for offline use. There are, however, better models that are specifically tailored to increase German performance. For example, you could support this prime line model. https://huggingface.co/primeline/parakeet-primeline @DevEmperor : I know you are German, so: auf geht's! : )

[MyButtermilk]

https://huggingface.co/geier/deskscribe-parakeet-primeline-onnx

[DevEmperor]

Shipped! 🎉 The Parakeet German (primeline) model (primeline/parakeet-primeline) is now available as an on-device (offline) transcription model.

How to use it: Settings → Transcription → on-device models → download Parakeet German (primeline) (~670 MB), then select it. No audio ever leaves the device. Attribution (CC BY 4.0) is listed under Settings → Language Data Attributions.

I converted the .nemo checkpoint to the sherpa-onnx ONNX int8 format (encoder/decoder/joiner + tokens) and mirror it on the project's own release for a stable source.

Small post-launch note in case it helps anyone doing NeMo→sherpa-onnx exports: the first build crashed on-device because sherpa-onnx decides whether a NeMo transducer is a TDT model purely by checking for the substring tdt in the encoder's url ONNX metadata. Since primeline's HF URL doesn't contain "tdt", the runtime treated it as a plain RNNT and aborted on the joiner output-size check (the 5 extra outputs are the TDT duration bins). Fixed by tagging the encoder metadata accordingly — no re-conversion needed.

Thanks for the suggestion, @MyButtermilk!