How To Control AI Voiceover Pronunciation for Names and Jargon With Custom Dictionaries

The CEO's name is Siobhan. The product is called NGEN-X. The patient's medication is adalimumab. Your AI voiceover just pronounced all three of them wrong, and the executive team heard the demo.

Mispronunciation is the fastest way to lose credibility in any AI-generated audio. A training video that stumbles over the company founder's name. A dubbed broadcast that mangles a city name. A product demo where the AI reads an acronym letter-by-letter when the audience says it as a word. AI voiceover pronunciation errors are not rare edge cases. They happen every time a text-to-speech system encounters a word that was not well-represented in its training data, which includes most proper nouns, brand names, medical terminology, legal vocabulary, and industry acronyms.

Custom dictionaries fix the problem at the source. Rather than editing audio after the fact or rewriting scripts to avoid difficult words, a pronunciation dictionary tells the TTS engine exactly how to say a specific term, every time, across every language and every project.

Why AI TTS Pronunciation Fails on Specialized Terms

Text-to-speech models convert written text into speech using a component called grapheme-to-phoneme (G2P) conversion. The model looks at the letters in a word and predicts the corresponding sounds. For common English words, the prediction is accurate. For everything else, it is a guess.

The problem is structural, not a bug. G2P models learn from training data that skews toward common vocabulary. Proper nouns appear less frequently and in less consistent patterns. "Siobhan" follows Irish phonetic rules that English G2P models do not know. "NGEN-X" could be an acronym (pronounced letter-by-letter) or a brand word (pronounced as "Engine-X"), and the model has no way to tell which one you intended.

Three categories of words consistently trip up AI voiceover pronunciation:

• Proper nouns: personal names, place names, company names, brand names
• Acronyms and initialisms: some are spoken as words (NASA, SCUBA), others as letters (FBI, API), and many are ambiguous (SQL, GIF)
• Domain-specific jargon: medical terms (adalimumab), legal terms (certiorari), technical terms (kubectl), and any vocabulary specific to your industry

What Are Custom Dictionaries?

A custom dictionary is a lookup table that maps specific written terms to their correct pronunciations. When the ​text-to-speech engine encounters a word that appears in the dictionary, the dictionary entry overrides the model's default G2P prediction.

CAMB.AI calls this feature Dictionaries. You define the correct pronunciation for any term, and that pronunciation applies consistently across all projects, all languages, and all voices. The dictionary entry travels with your content, so a term that was corrected once stays corrected everywhere.

How Dictionaries Work in Practice

Setting up a dictionary entry is straightforward. You provide the written form of the word (how it appears in the script) and the spoken form (how it should sound). For example:

• Written: "NGEN-X" → Spoken: "Engine X"
• Written: "Siobhan" → Spoken: "Shih-vawn"
• Written: "kubectl" → Spoken: "cube control"
• Written: "GIF" → Spoken: "jif" (or "gif," depending on your house style)

The system uses the spoken form whenever the written form appears in any script processed through ​DubStudio or the CAMB.AI API. No re-recording. No script rewriting. No per-project corrections.

Common AI Voiceover Pronunciation Problems and How To Fix Them

Different types of pronunciation errors require different dictionary approaches. Here are the patterns that production teams encounter most frequently.

Personal and Place Names

Names are the single largest source of text-to-speech pronunciation errors. A name like "Nguyen" (Vietnamese), "Bhattacharya" (Bengali), or "Xiaochen" (Mandarin) follows phonetic rules that an English-trained model does not recognize. Place names like "Worcestershire," "Louisville," or "Reykjavik" are equally unpredictable.

The fix: add every name that matters to your Dictionaries. For content that features recurring speakers, interviewees, or locations, build the dictionary before production begins. The time investment is minimal, and the payoff is immediate.

Acronyms and Abbreviations

The challenge with acronyms is ambiguity. "AI" is almost always said as two letters. "API" is always three letters. But "SQL" could be "sequel" or "S-Q-L" depending on your audience. "AWS" is always a word, but "WYSIWYG" is always a word.

The fix: define every acronym in your Dictionaries with the pronunciation your audience expects. CAMB.AI's Dictionaries apply globally across your ​AI translation and dubbing pipeline, so the same acronym is handled consistently whether the output is English, Spanish, or Japanese.

Numbers, Dates, and Currencies

TTS models handle most standard numbers correctly, but edge cases appear with phone numbers, model numbers, version numbers, and financial figures. Currency symbols combined with abbreviations (like "€1.5bn") are particularly worth defining in your dictionary.

Foreign Words in Monolingual Scripts

An English script mentioning "schadenfreude," "coup d'état," or "kaizen" creates a language-switching challenge. Dictionary entries for loanwords ensure the ​voice synthesis model pronounces each term according to the source language's rules.

Building a Pronunciation Dictionary for Your Organization

A well-maintained dictionary becomes a reusable asset. Start by auditing existing voiceovers and noting every mispronunciation. Categorize the errors: names, acronyms, jargon, or formatting issues.

Prioritize high-frequency terms. A CEO's name matters more than a one-time reference. A product name appearing in every demo matters more than a competitor mentioned once.

Test before publishing. Use the preview function in DubStudio to listen to dictionary entries before they go live. A phonetic spelling that looks right on paper may not sound right when synthesized.

Integrate with your ​voice library. Dictionaries and voice profiles work together. A cloned voice combined with correct pronunciation produces output that sounds both like the right person and says every word correctly.

Stop Correcting the Same Pronunciation Twice

Every mispronunciation is a fixable problem, and custom Dictionaries ensure you only fix each one once. Whether your content features executive names, medical terminology, sports team names, or technical jargon, a well-built dictionary turns AI voiceover pronunciation from a recurring headache into a solved problem. Define it once, and your ​text-to-speech output gets it right every time, in every language, for every project.

Get started for free →