Why Custom Vocabulary Matters
Dictation tools are trained on general English, so anything outside that, a client's last name, a product codename, industry jargon like "arthroscopy" or "Kubernetes", gets mangled more often than common words. If you dictate the same 20 unusual terms every day, even a 90% accuracy rate on them means you're correcting the same mistakes constantly. Scrybapp handles this differently than Apple's built-in Dictation, worth understanding before you spend an hour trying to build a vocabulary list that barely helps.
What You'll Learn
- How Apple Dictation's vocabulary options actually work, and their limits
- Why Whisper-based models like Scrybapp already know more jargon than you'd expect
- How context improves accuracy for names and technical terms
- Practical tips for dictating uncommon words reliably
Method 1: Apple Dictation's Vocabulary Options
Setting It Up
macOS doesn't have a proper custom vocabulary list, there's no menu where you type in "Kowalczyk" or a product name and teach the system to recognize it going forward. The closest thing is the Text Replacement feature in System Settings > Keyboard, which maps a shortcut phrase to expanded text, but that's a fixed substitution trick, not vocabulary training. It works for a handful of terms you type the shortcut for, not for dictation accuracy in general.
Limitations
- No real training — Apple Dictation's underlying model doesn't adapt to your speech or vocabulary over time, so mistakes repeat indefinitely.
- Text Replacement isn't dictation-aware — it substitutes typed shortcuts, it doesn't help Dictation recognize a spoken name correctly in the first place.
- Proper nouns are the weak point — names, brands, and acronyms outside common English get autocorrected into the nearest dictionary word.
- No per-document context — Dictation treats every sentence in isolation, so it can't infer from surrounding text that an unusual phrase is a product name, not two random words.
Method 2: Using Scrybapp for Uncommon Words and Names
Why Scrybapp Is Ideal For This
Scrybapp runs on OpenAI's Whisper model, trained on a much broader dataset than Apple's on-device model, including technical writing, multiple languages, and a wide range of proper nouns. That means it recognizes terms like "PostgreSQL," "amoxicillin," or uncommon surnames out of the box far more often than Apple Dictation does, because the model has actually seen those words in training rather than pattern-matching them against a small dictionary.
- Works everywhere — runs as a menu bar app, types into any active text field, no special integration needed.
- AI-powered accuracy — the Whisper model handles accents, jargon, and complex sentences using broader language context, not fixed command lists.
- 100% private — all processing happens on-device, so even sensitive names and terms never leave your Mac. See HIPAA-compliant dictation for why that matters in medical and legal fields.
- One-time purchase — $19 once, no subscription.
Step-by-Step Setup
- Step 1: Download Scrybapp.
- Step 2: Grant microphone and accessibility permissions.
- Step 3: Choose your keyboard shortcut.
- Step 4: Open the target app, press the shortcut, and speak the uncommon term clearly at normal pace, slower unnatural pacing can actually confuse the model.
- Step 5: Press again to stop, review the output, and if a name is consistently wrong, say it once more with a clarifying phrase right after, like spelling it out letter by letter, which Whisper will often pick up correctly the second time.
Tips for Improving Accuracy on Specific Terms
A few habits help regardless of which tool you use. Say uncommon names in full sentences rather than in isolation, giving the model more context than saying the name alone. If a term is genuinely ambiguous, like a homophone of a common word, spell it out once and the model tends to lock onto the correct spelling for the rest of that recording. For repeated technical vocabulary in fields like medical dictation or software development, dictating full phrases rather than single words also improves recognition, since Whisper leans on sentence-level context more than word-by-word guessing.
None of this requires a training screen or a vocabulary list you have to maintain. The model already knows more than most people expect, and the accuracy gap that used to require manual correction on every uncommon name has mostly closed as the underlying models improved. If you dictate the same jargon daily and it's still coming out wrong more than once in ten tries, that's usually a background noise or microphone issue rather than a vocabulary gap, worth checking your microphone setup before anything else.