THE SHORT VERSION
Keep pronunciation rules versioned and model-specific. Test important terms inside real sentences rather than approving isolated words alone.
Repeatedly fixing the same product name is a sign that a narration project needs a pronunciation reference. ElevenLabs documents pronunciation dictionaries, with support that depends on the selected model and rule type. We checked that guide on 30 October 2026.
Gather the terms that matter
Start with names, technical vocabulary, abbreviations, and locations that appear in the approved script. Record the normal written form alongside the desired spoken form. Include language and a note about context when one spelling has several valid readings.
Ask an appropriate speaker or subject expert to approve critical names. An invented phonetic spelling is not an authoritative reference. If several people disagree, resolve the intended audience’s pronunciation before making a reusable rule.
Keep the list small enough to review. A dictionary full of speculative exceptions can introduce new mistakes that are difficult to trace.
Check compatibility first
Read the current provider guide for the chosen model. ElevenLabs distinguishes alias replacements and phonetic approaches and describes language and model limitations. Do not assume a rule that works with one model behaves identically with another.
For an API implementation, record the dictionary and version used with each generation. For a product interface, confirm the dictionary is actually attached to the project or generation route. API documentation describes an integration route; controls in the product interface can differ.
Use sentences as the test unit
Prepare short passages containing each important term at the beginning, middle, and end of a sentence where relevant. Include the surrounding words that will appear in the real narration. Review the full phrase, not just the corrected syllables.
Check that a replacement has not altered an ordinary word with the same spelling. Also inspect acronyms and plural forms. A broad substitution can solve one line while creating an unnatural reading elsewhere.
Keep the visible caption spelling separate from any speech-oriented alias. Viewers should see the approved brand or technical term rather than a workaround written only to influence pronunciation.
Version changes and retain accepted audio
Change one rule at a time and note which passage justified it. Save accepted samples with the script, voice, model, and dictionary version. If a later generation sounds different, that record gives you something specific to compare.
Recheck critical passages when changing a voice or model. Reuse is a convenience, not evidence that pronunciation remains correct forever. Record usage and correction attempts to budget recurring narration.
For the full sequence from script approval to final listening review, see our AI voiceover workflow.