Features

Automatic language detection always translates to English.

1y agoPending

I haven’t used Whisper in a while and thought I’d try it again. However, I’m having a hard time making it work with automatic language detection.

I switch between German and English input all the time and I need automatic detection. When I activate automatic detection, I always end up having German dictation translated to English. When I set it to either German or English, I always receive, well German or English.

I expect Automatic language to transcribe whatever input language to the same output (which I think used to work like that).

Aside: My user dictionary must have been wiped by a recent update (haven’t used SW in a while but never uninstalled it) Tested with Ultra Cloud, Ultra Local and Nova 3 Cloud.

and 1 other

3 comments

  1. Same for me.

    Platform: Linux (Omarchy, Arch-based, kernel 7.2.15), package superwhisper 1.2.0-1 (metadata appVersion: 0.1.0) Processing mode: Cloud (S1-Voice sv-1, S1-Language sl-1) Desktop locale: LANG=en_US.UTF-8

    What happens With Language set to Automatic, every dictation is sent as English, whatever language I speak. Each history entry recorded on Automatic has "languageSelected": "en" in its meta.json. It looks like Automatic resolves to the desktop locale before transcription instead of letting the model detect the language.

    - Swedish turns into nonsense English. "Jag vill prata svenska, förstår du?" came out as Joghwil Bratasvenska, Verstorle. - Spanish usually comes out right, but it is still tagged en, so the model seems to recover by luck. - If I set Language to Swedish by hand, the entries show "languageSelected": "sv" and Swedish transcribes correctly.

    Steps to reproduce 1. On Linux with an English desktop locale, set Language to Automatic (superwhisper general show shows "language": "auto"). 2. Dictate a sentence in Swedish, or any language other than the locale's. 3. Check superwhisper history list: languageSelected is "en" and the text is a garbled English attempt.

    Expected Automatic should let the model detect the spoken language, as it does on macOS, rather than fixing it to the desktop locale.

    Related - Automatic language detection always translates to English (https://superwhisper.userjot.com/board/p/automatic-language-…): similar symptom, but there the model guesses wrong. Here detection never seems to run. - A way to limit Automatic to a shortlist (for me English, Spanish, Swedish) would also make detection more reliable for multilingual users.

  2. vv tt11mo ago

    Same

  3. Yeah, I ended up having several modes for several languages, and it makes me super slow, let alone unable to dictate multilingual content, which works flawlessly in the GPT app.