Parakeet v3 often tries to transcribe in English instead of Portuguese

[dserodio]

First of all, thanks for making this awesome open source project, it's much appreciated.

I want to use it mostly to transcribe Portuguese, so I'm using the "Portuguese - Parakeet v3" model. However, it often tries to transcribe my speech to English, which fails badly. I'm not sure if it's completely a problem with the model or if something could be improved in how Hex prompts the model.

I understand how difficult it is to test languages other than your native one, but I'm more than happy to help test and troubleshoot this in any way I can.

[kitlangton]

Selecting Portuguese in HEX does not force Parakeet v3 to recognize Portuguese. I traced the setting through our native backend: HEX passes the language hint, but the v3 model we ship does not use it to guide decoding. It determines the language from the audio, as described in NVIDIA’s model documentation. Our picker should make that distinction clearer.

Could you try Portuguese → Whisper large-v3-turbo in the local transcription picker? Unlike Parakeet v3, Whisper uses the selected language to guide transcription. This is an alternative to test, not a guaranteed fix for your recordings.

To investigate the incorrect output itself, could you also share your HEX version, whether you speak Brazilian or European Portuguese, and whether OpenCode rewriting is enabled in the matching Mode? If you are comfortable sharing a short, non-sensitive recording that reproduces the problem, the expected Portuguese text and actual output would help us distinguish recognition errors from later rewriting. No private recordings or full logs are needed. Thanks for offering to help test.

[dserodio]

Thanks for the detailed answer.
I'm now using Hex 2.1.16 (I don't recall which version I was using when I opened this issue, but it was the most current version at that moment).
I speak Brazilian Portuguese, and I'm not using OpenCode rewriting. I'll report a recording reproducing the problem soon.