mkiol
STT language auto-detection
Currently, the STT Dbus API requires mentioning a model based on a language. However, Whisper has a model that does language autodetection.
It would be nice if one could call SttTranscribeFile (and related APIs, I suppose) with string:lang:auto, such that the language gets inferred by the STT. The relevant signals should then return the detected language (but I think it's mostly there already).
Additionally, the relevant user interfaces (I'm on SailfishOS) should get some element that allows downloading this general model.
Hi, thanks for the idea. Language auto detection can indeed be useful.
I think I can add this in the next version.
Automatic language detection has been implemented in v4.6.0 ~~(not yet uploaded to OpenRepos)~~.
To use it, you need to download a model from "Auto detected" language category and pass "auto" as a language in Speech Note API.
STT with "Auto detected" models are a bit slower than models with a defined language, so if you know the language, it is still better to use models for a specific language. In v4.6.0 the speed of STT with Whisper is much much quicker, so maybe it is not a problem.
Changes: 5d0b9655525bbeeeee70b134f530b767575f02c4 57e1ee3cf980b09597a8f7bacc600607c260b1c0
This is beautiful, thank you! <3
Untested patch for Whisperfish: https://gitlab.com/whisperfish/whisperfish/-/merge_requests/613