Teach AICHE the Words You Actually Use

Short answer: open AICHE settings, add up to 50 words you use often (coworker names, product names, acronyms, code identifiers, brand names with weird capitalization), optionally write a phonetic hint, and they get spelled correctly in every recording on every device.

The problem this solves

Voice transcription is good at common English and bad at the words specific to your life. Your CTO is named "Saoirse", your product is called "Kintsugi", your library is "rkyv", and the model keeps writing "Sersha", "Can Sugi", and "archive". You end up fixing the same five words by hand in every note, or you stop using voice for anything you'd actually send.

How it works

1. Open AICHE settings on any device. Custom Vocabulary lives in the same panel on iPhone, iPad, Android, macOS, Windows, and Linux.
2. Tap Add Entry and type the word as you want it spelled. Capitalization, punctuation, and case are preserved exactly: kubectl, rkyv, Kintsugi, Saoirse.
3. Optionally add a phonetic hint - how you actually say it out loud. "Saoirse" gets the hint sersha. "rkyv" gets archive. "kubectl" gets cube control or cube cuttle, whichever you say.
4. Save. The entry syncs through your end-to-end encrypted account to every other AICHE surface within seconds.
5. Press ⌃+⌥+R on Mac or Ctrl+Alt+R on Windows / Linux, speak a sentence containing the word, and AICHE inserts the correct spelling at your cursor. On iPhone, iPad, Apple Watch, and Android, tap the mic instead - same dictionary, same correction.
6. When you reach the 50-entry cap, prune the entries you no longer use. The dictionary is a working set, not an archive.

Cross-platform sync

The dictionary is part of your AICHE account, not a per-device setting. Add an entry on your iPhone while waiting for coffee, and by the time you sit at your Linux desktop the same word is in the dictionary there. The same applies to iPad, Apple Watch (which pulls dictionary updates over its own connection), Android, macOS, Windows, the Chrome extension, and the Obsidian plugin.

Sync runs over the same end-to-end encrypted channel as your notes (AES-256-GCM with an Argon2id-derived key, passphrase set by you). We can't read your dictionary on our servers. Neither can anyone else.

This matters more than it sounds. Most voice apps that offer a custom dictionary store it on the device where you typed it. Switch laptops, lose your words. Reinstall, lose your words. AICHE treats the dictionary as account data, so the words you taught it once stay taught everywhere.

Applies to every transcription mode

Custom Vocabulary runs as part of the post-Whisper polish pipeline, before the text reaches you. That means it applies the same way regardless of what you're doing:

• Hotkey dictation into any desktop app (editor, email, Slack, terminal, IDE).
• Mobile capture on iPhone, iPad, and Android - voice notes, journals, meeting recordings.
• Apple Watch recordings, which upload directly from your wrist over Wi-Fi, cellular, or the paired iPhone tether - transcripts appear on your iPhone via cloud sync.
• Chrome extension inserts into web text fields.
• Obsidian plugin voice notes inside your vault.
• REST API calls on the Pro tier - the same vocabulary your interactive app uses applies to programmatic transcription.
• Auto-translation to English - if you speak in another language and ship clean English, the dictionary still enforces your proper-noun spellings on the English output.

You don't pick which surfaces get the dictionary. You taught it once. It's everywhere.

For developers

A 50-entry dictionary covers more of a working codebase than it sounds like. Realistic entries for a backend engineer:

• Repo and service names (auth-gateway, Kintsugi, Foundry)
• Library and tool names that Whisper mishears (rkyv, tokio, nginx, kubectl, pnpm)
• Internal acronyms (SSO, PII, RBAC) where you want the exact case
• Coworker names and Slack handles
• Brand names with non-standard capitalization (iCloud, PostgreSQL, OAuth)

Pair this with the Pro Software Development profile, which tunes the recognition pass itself for code-adjacent speech (kebab-case flags, snake_case identifiers, common library names). The profile changes what the model hears. The dictionary corrects what comes out. Together they catch most of the cases where raw dictation fails for engineering work.

For the 50-entry cap: prune. Once a project moves out of your active rotation, replace those entries with the words from the project you're in this week. The dictionary is meant to track your current vocabulary, not every word you've ever used.

Tips

Use phonetic hints for anything you don't say the way it's written. Whisper guesses spelling from sound. If the sound and the spelling disagree (Saoirse, rkyv, tmux, NGINX), the hint is the bridge.

Let AICHE write the hint for you. Tap the mic in the sounds-like field and pronounce the word out loud - AICHE transcribes what it heard and auto-fills the phonetic hint. Faster than guessing the spelling-of-the-sound yourself, and it captures exactly how you say it.

Add words as they bite you. The fastest way to fill the 50 slots well is to keep the dictionary panel one tap away and add a word the first time you see it transcribed wrong. Don't try to brainstorm the list cold.

Capitalization sticks. If you write iCloud, AICHE outputs iCloud, not "Icloud" or "I Cloud". The same goes for OAuth, JavaScript, PostgreSQL, and every brand with intentional case.

Combine with the Software Development profile (Pro). Profile tunes the recognition. Dictionary corrects the output. The two layers compound.

Don't try to teach it sentences or grammar. Custom Vocabulary is for nouns, names, and identifiers. Cleanup, punctuation, and filler removal are handled by the polish pipeline on every recording.

Result: the five words you used to fix by hand on every voice note get spelled correctly the first time, on every device, with no extra step.

Works With

Custom vocabulary applies to every recording, but it makes the biggest difference when names matter:

• Slack - teammate names and project names in messages
• Claude Code, Cursor - repo names and service names in prompts
• VS Code - library names and internal tools in documentation
• Notion - project names and client names in databases
• GitHub - PR descriptions with correct internal terminology

Try it now: open AICHE settings, add three words you've watched the transcript butcher this week (a coworker name, a product name, a tool you use daily), give each one a phonetic hint, then dictate a sentence using all three and watch them land right.