feat: add custom transcription prompt setting for Whisper models

[egsok]

Problem

Whisper often drops punctuation entirely, producing a wall of unformatted text — especially for non-English languages. This is a well-known issue in the community:

• openai/whisper#557 openai/whisper#194 — punctuation loss discussions
• OpenAI Whisper Prompting Guide — official guidance on using initial_prompt to steer style

The documented solution: pass a well-punctuated paragraph as initial_prompt. Whisper doesn't follow instructions — it copies the style of the prompt. A paragraph full of commas, question marks, and em-dashes nudges the decoder to keep producing punctuation.

What I tested

I tested this in my fork with a hardcoded Russian punctuation prompt, and it works remarkably well. Russian transcriptions went from completely unpunctuated maybe 5% of the time to basically always properly formatted output with commas, periods, question marks, em-dashes, and quotation marks — consistently, across different recording lengths.

Why not just hardcode a default prompt

The prompt also acts as a language hint. A hardcoded prompt in one language biases Whisper's language detector — e.g., an English prompt causes Russian speech to be transcribed as English when the language is set to "Auto". So pre-filling a default prompt would break the experience for anyone using auto-detection with a non-English language.

Worth noting: a prompt in one language doesn't prevent Whisper from recognizing words in another. For example, a Russian prompt with language set to "Auto" works fine for mixed Russian/English speech — English words are still transcribed correctly.

In the future, we could consider auto-populating the prompt based on the selected transcription language — but for now, empty by default is the safe choice.

Solution

A new Transcription Prompt setting under Settings → Advanced that lets users provide a sample text to guide Whisper's output style.

Key design decisions:

• Empty by default — no language bias introduced for users who don't need it. When Transcription Language is set to "Auto", a non-empty prompt in a specific language can reduce language detection accuracy, so opting in is intentional.
• 10-language preset dropdown (EN, ES, FR, DE, PT, IT, RU, JA, ZH-CN, ZH-TW) — each preset uses native punctuation conventions (Russian «ёлочки», German „Gänsefüßchen", French « guillemets », Japanese 「括弧」, Chinese ""引号""). Users can also write their own prompt.
• Token-aware budget — Whisper's initial_prompt window is 224 tokens. Custom Words are prepended and share that budget, so the prompt is capped at 112 estimated tokens (half). A per-script token estimator (CJK ~2.2 tok/char, Cyrillic ~0.5, Latin ~0.25) enforces the limit in real time. A progress bar with color coding (gray → yellow at 80% → red at 95%) replaces the old character counter. A hint below the bar explains the shared budget: "a shorter prompt leaves more room for custom words."
• Whisper-only — this setting only affects Whisper models. When a non-Whisper model is selected (Parakeet, GigaAM, Moonshine, Canary, Cohere, SenseVoice), a warning is shown. The setting remains visible and persisted so users don't lose their prompt when switching models.

Changes

Backend (Rust):

• settings.rs — new transcription_prompt: Option<String> field
• transcription.rs — concatenate prompt after custom words into initial_prompt. Custom prompt is placed last — Whisper truncates from the left, so dictionary words (lower priority) get truncated first
• shortcut/mod.rs — new update_transcription_prompt command
• lib.rs — register the command

Frontend (TypeScript/React):

• New TranscriptionPrompt.tsx component with preset dropdown, textarea, token-aware progress bar, and contextual warnings
• AdvancedSettings.tsx — include the new component
• settingsStore.ts — wire up the setting to the backend command
• bindings.ts — add transcription_prompt to AppSettings and updateTranscriptionPrompt command
• en/translation.json — all user-facing strings

Test plan

• Select a Whisper model → setting is shown without warnings
• Select a non-Whisper model → yellow "Whisper only" warning appears below
• Pick a language preset → textarea populates with sample text
• Switch preset to "None" → textarea clears, setting is saved as null
• Type a custom prompt → saved on blur, persists across app restart
• Set Transcription Language to "Auto" with a non-empty prompt → language detection warning appears
• Set Transcription Language to a specific language → no warning
• Transcribe with Russian preset on Whisper Large → punctuation (commas, periods, em-dashes, «quotes») present in output
• Transcribe with empty prompt → default Whisper behavior (may lack punctuation)
• Verify Custom Words still work alongside the prompt
• Token budget hint appears when prompt is non-empty
• Progress bar fills slowly for Latin text (~0.25 tok/char), fast for CJK (~2.2 tok/char)
• At 100% budget → typing blocked; deleting text → bar decreases, typing resumes
• Each preset fills the bar to ≤69% (leaving room for Custom Words)
• Bar turns yellow at 80%, red at 95%

[cjpais]

I think this is probably a good idea. I'm just not 100% sure about the UI for this yet.