Remote Control dictation is unusable for non-native / mixed-language speech, and there is no way to plug our own STT

I run a robotics workshop and drive Claude Code from my phone and laptop through Remote Control, by voice, in French mixed with English technical terms, for hours a day. On the rig itself I use local Whisper and it is fine. In Remote Control I have to use the built-in dictation, and it is the weakest part of the whole product.

Examples from today, all in one afternoon:
- "leur STT" became "le reste été"
- "Whisper Flow" became "Wishpower" and "Wishport"
- "ElRobot" (a product name) became "elle robote"
- "du matos, on a déjà un bras" came out as "Matosse on a déjà presque sur peut-être un braille"
Product names, acronyms and code identifiers get rewritten into unrelated French words several times per message. Every message needs a re-read and often a re-dictation. This is the door to the product: whatever is behind it arrives damaged.

Two things would fix it, either one:
1. A dictation engine that handles mixed French/English and technical vocabulary at least as well as Whisper large does (the CLI dictation page says it is tuned for coding vocabulary; the Remote Control one clearly is not).
2. An open door: let us choose our own STT provider or key (Whisper, OpenAI, Deepgram, local) for dictation in the app and in Remote Control. Then you do not have to win at STT, you just have to not block us.

Until then the honest option is to ship no dictation button in Remote Control rather than one that costs trust on every message. I am a paying Max subscriber and this single detail is the one that makes me look at the competition.

Environment: Remote Control from iPhone and macOS Chrome; session on Linux (Pop!_OS), Claude Code 2.1.263, language set to French.

Start with the Remote Control dictation entry point and compare it with the CLI dictation page, which the issue says is tuned for coding vocabulary. Reproduce mixed French/English speech and technical terms on iPhone or macOS Chrome with a Linux session, then establish whether the scope is improved accuracy, user-selected STT providers, or both, and validate the resulting dictation flow.

由索引模型根据 Issue 内容生成。