Every TTS user hits the same wall: the voice is great until it meets your brand name, your CEO's surname, or that industry acronym — and mangles it. One mispronounced word repeated across a hundred videos quietly erodes your credibility.
A TTS pronunciation dictionary fixes this systematically: a maintained list of problem words with their correct phonetic renderings, applied to every generation. This guide shows how to build one that actually works.
Key Takeaways
- Collect mispronunciations as you find them — the dictionary grows from real errors, not theory.
- Phonetic respelling ('say it like it's spelled') fixes 80% of problems without touching SSML.
- SSML phoneme tags give precise control for the stubborn 20%: names, loanwords, and acronyms.
- Test every entry in context — a pronunciation that's right in isolation can sound wrong in a sentence.
- Version your dictionary and re-test after voice model updates; new models pronounce differently.
Why TTS Mispronounces Your Words
TTS models learn pronunciation from massive text datasets, which means they handle common words brilliantly and unusual words statistically. Your brand name, invented product lines, surnames, and industry jargon are all 'unusual' by training-data standards — so the model guesses, and guesses are often wrong.
Acronyms are the classic trap: is 'GIF' hard-G or soft-G? Is 'SQL' 'sequel' or 'S-Q-L'? Your organization has a house answer; the TTS model has a statistical one. Without a dictionary, you get the model's answer — inconsistently, across every video.
Loanwords and names cross language boundaries unpredictably. A French brand name in English narration, a German surname in a Spanish video — each needs an explicit decision. The dictionary is where those decisions live, so every future generation inherits them instead of re-guessing.
Building the Dictionary: Methods
Three methods, escalating precision:
| Method | How it works | Precision | Best for |
|---|---|---|---|
| Phonetic respelling | Write the word as it sounds | Good | Most brand/term fixes |
| SSML phoneme tags | IPA or X-SAMPA pronunciation | Exact | Names, critical terms |
| Alias/substitution | Replace with spoken equivalent | Perfect | Acronyms, symbols |
| Custom lexicon upload | Platform-level word list | Exact, global | Enterprise scale |
Start with phonetic respelling — it's fast and tool-agnostic. 'Qyura' mispronounced? Try 'kyoo-rah' or 'KEW-ruh' until the output sounds right. Keep a log of what worked; respellings are surprisingly non-obvious and you'll forget the winning variant.
Reserve SSML phoneme tags for words where respelling fails or precision is contractual — executive names, legal terms, regulated product names. The IPA learning curve pays off for a small set of high-stakes words.
Phonetic Spelling Techniques That Work
Think in syllables and stress. English TTS follows stress cues, so mark them: 'IN-voice' vs 'in-VOICE' produce audibly different results. Hyphens between syllables help the model parse unusual words; capitalization can nudge stress placement.
For acronyms, decide the house style first: spelled out ('F-B-I'), spoken as a word ('NASA'), or expanded ('Federal Bureau of Investigation' on first mention, 'F-B-I' after). Document the decision per acronym — inconsistency across videos looks sloppy.
Numbers and symbols need rules too: does '$1M' read as 'one million dollars'? Does 'v2.4' read as 'version two point four'? These seem trivial until a product launch video says them wrong. Add formatting rules to the dictionary alongside pronunciations.
Maintaining the Dictionary Over Time
A dictionary is a living document. Assign an owner — usually whoever produces the most voiceover content — and a simple intake: anyone who hears a mispronunciation logs the word, the context, and the desired pronunciation. Review the log monthly.
Re-test after voice model updates. New TTS models pronounce differently — sometimes better, sometimes newly wrong on words that used to work. A quick regression pass over your dictionary's entries after any platform update catches surprises before they reach published content.
Share it across the organization. Marketing, L&D, product, and support all generate voiceover now; one shared dictionary keeps the brand name pronounced identically everywhere. Store it where the content teams already work — a simple spreadsheet beats a sophisticated system nobody opens.
Frequently Asked Questions
What is a TTS pronunciation dictionary?
A maintained list of words your text-to-speech system mispronounces, each paired with its correct phonetic rendering — via respelling, SSML phoneme tags, or substitution rules. Applied to every generation, it ensures brand names, terms, and acronyms sound right consistently.
How do I fix a brand name TTS keeps mispronouncing?
Start with phonetic respelling: write the name as it sounds ('kyoo-rah' for Qyura) and test variants until the output is right. For stubborn cases, use SSML phoneme tags with IPA notation for exact control. Log the winning variant in your dictionary.
Should acronyms be spelled out or spoken in TTS?
Decide per acronym and document it: spelled out ('F-B-I'), spoken as a word ('NASA'), or expanded on first mention. Inconsistency across videos looks unprofessional, so the dictionary should record the house style for every acronym you use.
Do I need to learn IPA for a pronunciation dictionary?
Not for most entries — phonetic respelling handles about 80% of problems. Learn basic IPA only for the high-stakes words (executive names, legal terms) where respelling isn't precise enough and SSML phoneme tags are warranted.
How often should I update the pronunciation dictionary?
Review the mispronunciation log monthly, and run a regression test over all entries after any TTS voice model update — new models pronounce differently and can break previously working entries. Assign one owner to keep the process alive.
Related Guides
- Text-to-Speech API Guide for Developers
- How to Write Scripts for AI Voiceovers
- AI Voice Credits Explained
If this guide helped you plan your next project, put it into action right now — Try SynVoice free and generate your first AI voiceover in under a minute, no credit card required.