Local STT (Argmax sidecar) ignores spoken_languages — no language param in WebSocket URL

[adhdvital]

Bug Description

When using local STT (Argmax WhisperKit sidecar), Char does not pass the language parameter to the WebSocket connection URL. This causes Whisper to auto-detect language instead of using the user's configured spoken_languages setting.

Steps to Reproduce

1. Set Settings > Language & Vocabulary > Spoken languages: Ukrainian (uk)
2. Start a recording session with Ukrainian speech
3. Observe transcript is in English (translated from Ukrainian instead of transcribed)

Root Cause

The WebSocket connection in the Argmax adapter (crates/owhisper-client/src/adapter/argmax/live.rs) connects to:

ws://localhost:{PORT}/v1/listen

without appending ?language=uk or any language parameter.

The sidecar (WhisperKit v1.11.4) does accept the language query parameter (confirmed by manual WebSocket connection test with ?language=uk), but Char never passes it.

Evidence from Logs

ws_client::retry: connect_async: ws://localhost:54884/v1/listen

No query parameters present.

Expected Behavior

The WebSocket URL should include the configured language:

ws://localhost:{PORT}/v1/listen?language=uk

Additional Context

• Char version: Nightly (com.hyprnote.nightly)
• Model: openai_whisper-large-v3-v20240930_626MB
• macOS Sequoia
• spoken_languages: ["uk"] and ai_language: "uk" are correctly saved in settings.json but never read by the Argmax adapter
• Related issues: Language setting reverts to Polish after tab navigation #4570 (language setting reverts), Language mismatch guidance #3133 (language mismatch guidance)

Suggested Fix

In crates/owhisper-client/src/adapter/argmax/live.rs, read the spoken_languages setting and append it as a query parameter to the WebSocket URL when connecting to the local sidecar.

[goranmoomin]

We do send the user's primary language (that is supported by the model) over to argmax (via ArgmaxLanguageStrategy, if you are interested), and argmax does natively support multiple languages (i.e. below is something that the argmax with model parakeet-v3 has produced as the transcript)

The reason it's not in the logs is because the connection URI is stripped of info (for privacy reasons).

The primary language is determined by the order of the languages in your Transcription Spoken Languages:

If Ukranian is not the first, you might also want to update that so that Ukranian comes first (for the Whisper model, it's going to be the first language, for the Parakeet v3 model, it's going to be the first language that the model supports).

If that doesn't work, you might want to test with different models, if you still have issues feel free to ping us again!

[goranmoomin]

Actually just realized that this might be #4676 as well – if you're using Ukrainian as your primary macOS language, and seeing English as the summary language, we have a fix that will be included in the next release.