Custom Vocabulary: Stop Correcting Your Own Name

Custom vocabulary lists fix the #1 frustration with voice dictation: mangled names, acronyms, and domain-specific terms. By feeding a word list to the transcription engine, tools like Parrot ensure "Kubernetes" doesn't become "Cooper Netties" and your coworker's name comes through correctly every time. Here's how custom vocabulary works and how to set it up.

Why transcription engines struggle with names

Speech-to-text models are trained on general language. They optimize for the most probable sequence of words given the audio. For common English words, this works extremely well. For proper nouns, brand names, technical jargon, and abbreviations, the model has to guess - and it guesses wrong a lot.

This isn't a flaw in any specific provider. Whisper, Deepgram, and ElevenLabs all have the same issue. The model doesn't know that "Supabase" is a word, so it transcribes what sounds closest in its training data.

How custom vocabulary works

Custom vocabulary gives the transcription engine a hint: "these specific terms are likely to appear in this audio." When the model is deciding between "Cooper Netties" and "Kubernetes," the vocabulary list tips the scale toward the correct transcription.

In Parrot, you add terms in your profile settings. The list is stored locally in your SQLite database and sent alongside every transcription request. It works in all modes - local (Whisper.cpp), BYOK (your own API keys), and managed (we handle everything).

What to add to your vocabulary

Not every word needs to be in your vocabulary - common English words are already handled well. Focus on:

• People's names - coworkers, clients, doctors, anyone you mention regularly. Include full names and common shortened versions.
• Company and product names - your company, tools you use, products you discuss. "Figma," "Vercel," "Supabase," "Notion."
• Technical terms - domain-specific jargon. "GraphQL," "OAuth2," "PostgreSQL," "echocardiogram," "amortization."
• Acronyms and abbreviations - "YC," "SaaS," "HIPAA," "BP," "EHR." Include what they stand for if the model confuses them.
• Unusual spellings - brand names with non-standard capitalization or spelling. "macOS," "iPhone," "TypeScript."

How many terms do you need?

Start small. Add the 10–20 terms you use most frequently and that get transcribed incorrectly. You'll notice an immediate improvement. Over time, add terms as you encounter transcription errors - it's an iterative process.

There's no practical limit to the number of terms, but keeping the list focused is better than dumping in hundreds of words. A targeted list of 50–100 terms covers most people's needs.

Vocabulary + AI cleanup = near-perfect output

Custom vocabulary handles the transcription step. The AI cleanup step handles everything else - grammar, punctuation, filler words, formatting. Together, they produce output that reads like you typed it carefully, even though you were speaking stream-of-consciousness.

For example, you might say: "send the proposal to Priya Raghavan at Anthropic and CC the YC partners, um, make sure to mention the Series A timeline." With vocabulary and cleanup, the output is:

"Send the proposal to Priya Raghavan at Anthropic and CC the YC partners. Make sure to mention the Series A timeline."

Every proper noun correct. No filler words. Proper punctuation. Ready to send.

Getting started

In Parrot, go to your profile page and find the vocabulary section. Start typing terms and press enter to add each one. Your vocabulary syncs across all your dictation sessions immediately - no restart needed.

If you're dictating medical notes or legal documents, your vocabulary list will be longer and more specialized. That's fine - the more specific your list, the better the results. Join the waitlist to try it when Parrot launches.