Custom Dictionary for AI Transcription: Fix Names and Terms Automatically

Your CEO's name is not "Mark Suckerbird."

Your lead product is not "Slacks" or "Sales Horses" or "Jiro." And the client company you've been pitching for six weeks is definitely not "Acne Pharmaceuticals."

But if you've used AI transcription for more than a few meetings, you've seen exactly this kind of error—real names reduced to whatever the model thinks sounds closest. The meeting content is right. The participants are right. The timing is right. But every proper noun has been quietly mangled, and now someone has to go through the transcript and fix it before it goes into the meeting record.

This guide explains why AI transcription makes these mistakes, how a custom dictionary solves the problem at the source, and how to set one up so it works automatically on every recording you make.

---

Automate your meeting notes. MinuteKeep records your meeting and uses AI to transcribe, summarize, and extract action items. 9 languages, no subscription, 30 min free.

Why AI Transcription Gets Names and Terms Wrong

Whisper—the speech-to-text model behind most AI transcription tools, including MinuteKeep—is trained on hundreds of thousands of hours of general speech. That training gives it strong baseline accuracy across a wide range of accents, noise conditions, and languages. But "general speech" means exactly that: it reflects the distribution of words in everyday conversation, not the specific vocabulary of your organization or industry.

When the model hears a sound, it predicts the most statistically likely word. For common words, this works well. For words that rarely appear in training data—company-specific product names, uncommon surnames, niche technical terms—it picks whatever common word sounds closest.

This is why:

• "MinuteKeep" gets transcribed as "minute keep" or "Minute Cave"
• "Kubernetes" becomes "Cooper Nettie's" or "Cuber Netties"
• "Diaz" becomes "dias" or "Dietz"
• "EBITDA" becomes "EB ITDA" or sometimes just "ABIDA"
• "SaaS" gets written as "sauce" or "sass"

It's not random noise. The model is doing exactly what it was designed to do—it's just operating on statistics that don't include your world.

Benchmarking research confirms the pattern. Studies comparing Whisper against newer models specifically flag proper noun recognition as Whisper's weakest category. One comparative test found Whisper turbo had an 11% higher error rate than the top model specifically on proper nouns. OpenAI's own documentation notes the limitation and suggests prompt engineering as a partial workaround—but that approach requires technical access to the API and applies only at the point of transcription, not retroactively.

For professional use—where the meeting notes will be shared, acted on, or stored—these errors create real friction. Someone has to catch them. Someone has to fix them. And unless that person was in the meeting, they may not even know what the correct version looks like.

---

How a Custom Dictionary Solves This

A custom dictionary works differently from prompt-based approaches or model fine-tuning. Instead of trying to teach the AI before transcription, it applies a deterministic substitution layer after transcription, before the text reaches you.

The logic is simple: you define pairs of (wrong word → correct word). Every time the transcription engine produces the wrong word, the dictionary replaces it silently. You see the correct word in the output, every time, automatically.

This approach has a few practical advantages:

It's reliable. There's no probability involved. If the dictionary says "minute keep" → "MinuteKeep," that substitution happens without fail. You're not hoping the model gets it right—you're guaranteeing the output is correct.

It's cumulative. Every entry you add improves every future recording. You build the dictionary once and it keeps working.

It's immediate. No retraining, no model updates, no waiting period. Add an entry and it takes effect on the next transcription.

It handles what prompts can't. Prompt-based approaches require access to the API parameters, must be set up before each recording, and don't apply retroactively. A post-processing dictionary is set-and-forget.

---

MinuteKeep's Dictionary: How It Works

MinuteKeep's custom dictionary is built directly into the Settings screen under a dedicated Dictionary section. The setup is straightforward:

Step 1: Open the Dictionary screen. Navigate to Settings in the bottom tab bar, then tap Dictionary. You'll see a list of your existing entries (empty when you start) and an option to add new ones.

Step 2: Add a word pair. Tap the add button. You'll see two fields: one for the "wrong" word (what Whisper produces) and one for the "correct" word (what you want to see in the transcript). Fill both in and save.

Step 3: Test with your next recording. Make a recording that includes the term you added. After transcription, check the result—your correct version should appear automatically.

Step 4: Expand your dictionary as you go. After each meeting, scan the transcript for any remaining errors. Add the pairs you find. Over two or three meetings, your dictionary converges on your actual vocabulary and errors become rare.

The dictionary applies to every transcription from that point forward, across all nine languages MinuteKeep supports. Entries are stored locally on your device.

One practical note: when building your initial entries, it helps to work backward from real transcription output. Don't guess what Whisper will produce—record a short test clip, see what comes out, and add those specific "wrong" versions as your source terms.

---

Dictionary Entries by Industry

The following examples reflect common Whisper substitution errors in specialized fields. Use these as a starting point and adjust to your specific organization.

Technology

Wrong (Whisper output)	Correct
kubernetes	Kubernetes
dock her	Docker
sauce	SaaS
AP eye	API
devops	DevOps
sequel	SQL
type script	TypeScript

Medical / Healthcare

Wrong (Whisper output)	Correct
at a ral fib	atrial fib
tpa	tPA
him a glob in	hemoglobin
cat scan	CT scan
em are	EMR

Medical note: A custom dictionary reduces errors but does not eliminate the need for clinical review before documentation becomes part of a patient record.

Legal / Professional Services

Wrong (Whisper output)	Correct
in demnification	indemnification
IP O	IPO
non disclosure	NDA
statue	statute

Legal note: Homophones—"statue" for "statute," "compliment" for "complement"—are phonetically identical. Dictionary matching catches consistent Whisper substitutions, but flag homophone-sensitive clauses for manual review.

Finance

Wrong (Whisper output)	Correct
ABIDA	EBITDA
EB ITDA	EBITDA
capex	CapEx
opex	OpEx
a are	ARR
m double r	MRR
cac	CAC

---

How Other Apps Handle Custom Vocabulary

It's worth understanding how competing tools approach the same problem, because the implementation differences affect how useful the feature is in practice.

Otter.ai has a custom vocabulary feature available on all paid plans. On the Pro plan ($8.33/month billed annually), you can add up to 100 names and 100 other terms per user. On Business plans, team vocabulary extends to 800 names and 800 terms shared across the organization. The vocabulary is applied at the transcription stage, which means it influences how Otter interprets incoming audio rather than correcting output after the fact. This is technically more powerful but also means it only works for new recordings—it doesn't retroactively fix existing transcripts.

Fireflies.ai offers custom vocabulary on Pro and Business accounts. The feature is input into a settings panel and "trains" the system on your terms over time. Like Otter, it operates pre-transcription rather than post-processing. Fireflies also distinguishes between custom vocabulary (for recognition) and topic tracking (for flagging keywords in transcripts)—two different tools with overlapping but distinct purposes.

Notta supports custom vocabulary of up to 200 words on Pro and 1,000 words on Business plans. The feature helps particularly with Japanese and English specialized terms, reflecting Notta's strength in multilingual contexts.

The pattern across these tools is that custom vocabulary tends to be a paid-tier feature, capped by plan level, and tied to ongoing subscriptions. If you cancel, you lose the vocabulary library.

MinuteKeep's approach is different in two ways: the dictionary applies post-transcription (deterministic, not probabilistic), and it's stored locally on your device. There's no per-tier word limit and no risk of losing your entries if you don't renew a subscription—because there is no subscription.

---

Ready to Stop Fixing Names Manually?

MinuteKeep is a pay-per-use AI transcription app for iPhone with a built-in custom dictionary. 30 minutes free, no account required.

Download MinuteKeep on the App Store

---

Best Practices for Dictionary Management

A dictionary is only as good as the entries in it. A few practices that make it easier to maintain:

Build it from real errors, not assumptions. Run a few recordings first, note what Whisper actually produces, and add those specific strings. Whisper's substitution patterns are predictable but not always obvious in advance.

Include capitalization variants. If your product name is "WorkflowIQ," add both "workflow IQ" and "workflowiq" as source terms—Whisper's output varies.

Add acronyms as both spoken and written forms. If people say "S-P-A" aloud, add that version. If they say "spa" as a word, add that too.

Add new names before major meetings. If a new client name or product launch will come up, add it the day before. It takes 30 seconds and eliminates post-processing.

Review the dictionary periodically. Products get renamed. People leave. An outdated entry can introduce errors rather than fix them.

---

Frequently Asked Questions

Does the dictionary work across all languages? Yes. MinuteKeep's dictionary applies to transcription output regardless of which of the nine supported languages was used. This matters particularly for multilingual meetings where product names or people's names appear in a non-English context—they're just as likely to be misrendered.

What if Whisper produces multiple different wrong versions of the same word? You'll need a separate dictionary entry for each variation. For example, "Kubernetes" might come out as "Cooper Nettie's," "Cuber Netties," or "kubernetes" in different recordings. Add all three as source terms pointing to the same correct output.

Does the dictionary change the transcript retroactively? No. MinuteKeep's dictionary applies to new transcriptions from the point you add the entries. Existing saved notes require manual correction in the Result view.

How many entries can I add? There's no documented cap. Practically, 20–100 entries covers most recurring errors for a given team or role.

Can I use the dictionary for phrases, not just single words? Yes. Multi-word phrases work as source terms—useful for client names or product names that span two words.

---

Key Takeaways

• AI transcription models like Whisper are trained on general speech and systematically underperform on proper nouns, technical jargon, and organization-specific vocabulary.
• A post-processing custom dictionary solves this at the output level: deterministic, cumulative, and set-and-forget.
• MinuteKeep's dictionary is built into Settings → Dictionary, requires no subscription, and works across all nine supported languages.
• Build your dictionary from actual transcription output rather than guessing what the model will produce.
• Other tools (Otter.ai, Fireflies, Notta) offer similar vocabulary features, but most gate them behind paid tiers and subscription plans.
• Periodic maintenance—adding new terms, removing outdated ones—keeps the dictionary accurate as your vocabulary evolves.

For a broader look at how MinuteKeep compares to other transcription apps, see Best Meeting Transcription Apps for iPhone (2026). If you want to understand the underlying accuracy trade-offs between transcription modes, see our guide on AI Transcription Accuracy: What the Numbers Actually Mean. And for additional tactics to reduce errors before and during recording, see How to Improve AI Transcription Accuracy: 12 Practical Tips.

---