Write a short pronunciation specification
For an illustrative inventory-app demo, the risky terms are the product name, warehouse names, item codes, and abbreviations. Make a small list with the written form, intended spoken form, language, and a sentence in which it appears. Ask the product owner to confirm any name whose pronunciation is ambiguous.
Do not begin with the entire three-minute narration. A short sample containing the difficult terms reveals whether the chosen voice and model can deliver the material understandably. Keep the original customer-facing spelling separate from any temporary spelling used to guide speech.
| Written item | Decision to record | Test context |
|---|---|---|
| Product name | Approved spoken name | Opening introduction and closing sentence |
| API | Letters or a spoken word | A sentence about an integration |
| Item code A-12 | How letters and numbers are read | An inventory lookup instruction |
| A date | Language and intended interpretation | A delivery scheduling example |
Choose the voice and model before fine-tuning settings
ElevenLabs' product guide identifies voice and model selection as major influences on speech output. Available settings differ across models, and generation is not fully deterministic. Keep the voice, model, and test script recorded so you can make a useful comparison.
For the inventory demo, compare candidate voices using the same sentence and intended language. Listen for whether a new user can distinguish the product name from a common word. A lively delivery is less useful if the instruction becomes difficult to understand.
Use an alias or a supported phoneme rule
ElevenLabs pronunciation dictionaries support word substitutions through aliases and phonetic rules with supported alphabets. The documentation limits phoneme behavior to particular models and languages; unsupported models can skip those phoneme rules. Check the current compatibility guidance for your exact generation route.
Use an alias when substituting a familiar spelling or expanded phrase produces the intended sound. Use a phoneme rule when you have an approved pronunciation and the chosen model supports that approach. Keep the rule narrow enough that it does not unexpectedly change an ordinary word elsewhere in the script.
Test the dictionary in full sentences
The linked dictionary guide describes an API workflow; check whether your chosen editor or API route supports the dictionary and model you intend to use. Attach the intended dictionary through that supported setup, then generate the short sample again. Compare the opening, a middle instruction, and the final sentence. A name that sounds acceptable in isolation may blend awkwardly into the words around it.
Change one relevant input at a time. If you change the voice, model, wording, and dictionary together, you cannot clearly identify the useful adjustment. Keep one approved sample as a reference and note which dictionary version generated it, especially when several people prepare narration.
Review numbers and pacing as separate problems
ElevenLabs' best-practice guidance explains that punctuation and supported controls influence pauses, with important differences between models. Do not assume an SSML pause tag that works in one model is supported by another. Check the current guidance before adding markup throughout a script.
For the inventory demo, write out an ambiguous number or abbreviation deliberately and listen to the result. Confirm that a quantity, item code, and date retain their intended meanings. If a sentence is difficult to follow, try clearer wording before trying to solve every issue with a voice setting.
Approve the final audio in its real destination
Listen to the narration alongside the screen recording. Check that the spoken action matches what is visible and that a pause does not make the next instruction sound like part of the previous one. Review the final audio after any edit that changes the timing.
Save the approved script, dictionary revision, voice, and model with the delivered file. When the product renames a feature, rerun the affected pronunciation samples rather than assuming an old setting will preserve the result. ElevenLabs is useful for this workflow when the team can maintain an understandable, repeatable review process.
Sources and editorial method
Official sources linked in this guide were checked on October 27, 2026. Product features, interfaces, and plan limits can change. The examples are proposed workflows, not reports of a hands-on benchmark or customer results. Confirm the documentation for your product version before implementing a workflow.