[purinda]

This lets you set the language the speech model transcribes in, e.g. English only, so a multilingual model doesn't drift into another language now and then. It's built on #65, which has merged.

Where

• The Language menu in the menu bar (Mac) and the tray (Linux, Windows). It sits after Cleanup and lists the languages for the model you're using. A model that can't be told a language shows Automatic and nothing else, greyed out.
• The model's row in Settings › Models, in the same spot as Cohere's language picker.

Both menus write the same sttLanguage setting. Automatic (no setting) is the default for every model. The setting is shared across models, so if you switch to a model that doesn't have the chosen language, it shows as Automatic, and the choice comes back when you switch to a model that has it.

What each model offers

Model	Languages	Automatic
Qwen3-ASR 0.6B, 1.7B	30 (from the model's support_languages)	the model finds the language
Qwen3-ASR 0.6B Sinhala	31 (the 30 plus Sinhala)	the model finds the language
Whisper large-v3 turbo	100	the model finds the language
Cohere Transcribe	14	English (Cohere can't find the language itself, so the menu shows "Automatic (English)")
Parakeet TDT v2, v3	none (can't be told one)	the model finds the language

English comes first in each list, then the rest by name.

Qwen3-ASR is kept to the language's letters

Testing showed telling Qwen3-ASR a language only replaces its own guess at it: with Sinhala chosen, English speech came out in English, and with English chosen, Sinhala came out in Sinhala. So when a language is chosen, Qwen3-ASR may now write only that language's letters, plus Latin letters for English words (as the Sinhala model writes them), and the digits, punctuation and spaces every language shares.

• Each Qwen language in speech-models.json names its script (ISO 15924: Latn, Sinh, Hani…).
• At each step the model picks the likeliest token whose bytes keep to those letters. A character split across tokens is tracked, and the reply can only end after a whole character.
• Mac: Qwen3-ASR is decoded in Qwen3ASR+ScriptLock.swift, on mlx-audio-swift's public model API, so there's no fork. Other models still use mlx-audio-swift's generate.
• Linux/Windows: the greedy decoder takes the same lock (script_lock.rs, qwen3_asr/locked_tokens.rs).
• Automatic never locks.

What it does (Mac Bench, default Sinhala Qwen model):

• On the 65 English eval clips, English locked gives exactly the same output as Automatic (WER vs said 12.8% both; p50 140 vs 158 ms).
• With English locked, a Chinese clip comes out as an English translation ("We'll meet you at 10:00 in the office.").
• It can't tell English from German, since both use Latin letters.
• With Sinhala chosen, other languages come out as Sinhala-script gibberish that can run to the output limit. This happened before the lock too; the loop guard misses multi-token Sinhala words.

How the language is told

• Qwen3-ASR is told the language by name, by adding language English<asr_text> after the assistant turn starts, as mlx-audio-swift's buildPromptText does.
• Whisper and Cohere are told the code, as Cohere was before.
• --language automatic clears the setting for a single run on Linux and Windows.

Not tested yet

• Whisper's 100 codes haven't been checked against a real Whisper tokenizer.
• The Linux/Windows tray, settings window and dictation code is behind cfg(linux|windows), so CI is the first place it compiles.
• The Linux OpenVINO lock is covered by unit tests only; I haven't run it on a real model.

Checks

• make test: everything passes. The two wall-clock cleanup timing tests failed once under load and passed on a rerun. The Mac app builds.
• cargo fmt --check, cargo clippy --workspace --all-targets -- -D warnings and cargo test --workspace pass on macOS.

🤖 Generated with Claude Code