Pronunciation Dictionary for AI Voiceover

Mispronounce the brand once and the whole spot is unusable

The most embarrassing AI voiceover failure isn't the wrong tone — it's getting your product name wrong. nginx read letter by letter, SaaS spelled out, a founder's surname mangled. In a long script those errors repeat dozens of times.

A pronunciation dictionary fixes exactly that: tell the model how a word should be read, and every generation obeys.

Real A/B

The same English line, before and after applying four rules:

Without the dictionary

With the dictionary

Text: Our SaaS runs on nginx. Ship a GIF with VoiceSmiths.

Rules:
  SaaS       → sass
  nginx      → engine ex
  GIF        → jif
  VoiceSmiths → Voice Smiths

Writing rules

A dictionary is a written-form → spoken-form replacement table. No phonetic alphabet required — just spell it how you want it said:

| Case | Written | Read as | | --- | --- | --- | | Acronym | SaaS | sass | | Tech term | nginx | engine ex | | Brand | VoiceSmiths | Voice Smiths | | Surname | Ng | Eng | | Ambiguous word | live (adj.) | lyve |

How to use it

1. Studio → Pronunciation: name the dictionary, add rules, create
2. Back on the voiceover tab, dictionaries appear as chips — tick the ones to apply
3. Stack several at once (e.g. "Brand terms" + "People"), applied in order

It matters most for long-form audiobooks: a proper noun appearing a hundred times gets fixed once.

Versus emotion tags

• Emotion tags like [whispers] control how it's acted
• The dictionary controls how it's pronounced

Use both together — they don't conflict, and neither leaks into subtitles.

Create my first dictionary →