Use a small, versioned pronunciation dictionary for repeated vocabulary. Resolve ambiguous numbers in the script, verify rule compatibility for the selected model, and approve every final audio asset by listening.
Pagember is an independent affiliate of ElevenLabs and may receive compensation for referrals.
A narration can sound polished and still get the most important word wrong. A product name pronounced two different ways, an acronym read as a word, or a date interpreted in the wrong order can distract the listener or change the instruction. ElevenLabs pronunciation dictionaries give a production team a reusable place to manage recurring terms instead of repairing each script independently.
The useful workflow starts with the intended spoken meaning, then chooses a supported rule and checks the result in context. This guide follows a hypothetical software company producing short website tutorials. It is researched guidance based on official ElevenLabs documentation reviewed on 14 October 2026, not a hands-on pronunciation benchmark. The example rules require listening review with your chosen voice and model.
Create a pronunciation register before a dictionary
Collect the terms that recur across the actual content library. Include product names, staff or presenter names used with appropriate permission, acronyms and technical terms. Give each entry an approved spoken form, language, owner and example sentence. Do not start by rewriting every ordinary word the model already handles well.
For the hypothetical company, the first register contains its invented product name, “API,” “CSV” and one internal feature abbreviation. The product owner approves the brand pronunciation. The technical editor decides whether the audience needs the acronym expanded on first mention. A language reviewer checks terminology in a localized tutorial.
| Written item | Decision to record | Review context |
|---|---|---|
| Invented product name | Owner-approved pronunciation | A short introduction and a complete instruction |
| API | Letters or an expanded first mention | A technical tutorial for its intended audience |
| 12.50 | Amount, measurement or version? | The surrounding sentence and unit |
| 06/07 | Which date, in which market? | The approved script’s explicit date |
This register is an editorial asset, not only an audio setting. It helps keep captions, transcripts and spoken instructions consistent. Preserve the normal written spelling where readers expect it; pronunciation work should not make the public transcript difficult to read.
Choose an alias or a phoneme rule for the actual model
ElevenLabs distinguishes alias rules, which substitute another text form, from phoneme rules, which specify pronunciation through a supported phonetic alphabet. Model support differs. Its current guide includes a model-by-model pronunciation matrix, so check the exact model you will use rather than copying an older tutorial’s universal rule. See the official pronunciation dictionary guide.
An alias is a practical first attempt when familiar text produces the desired sound. An acronym that should be spoken as separate letters may be tested with spaced letters. A full expansion can be clearer for a nontechnical audience. A phoneme rule may be useful when an ordinary respelling is unreliable, but someone still needs to verify the transcription and final audio.
Do not assume that inline SSML, inline phonetic text and dictionary phonemes are interchangeable. The supported technique depends on the model and surface. If a rule appears to be ignored, check compatibility before changing voices or generating the entire project again.
For a person’s name, use an approved pronunciation from that person where appropriate. Do not guess from their nationality or attempt to imitate their voice without authorization. A pronunciation dictionary changes how text is spoken; it does not provide permission to clone or impersonate a speaker.
A quick rule-selection path
- A familiar respelling works: try an alias inside a complete instruction.
- A precise recurring sound: check dictionary-phoneme support for the chosen model.
- One phrase needs repair: check a supported inline method before creating a broad replacement.
The compatibility table distinguishes dictionary rules, inline IPA and SSML. Choose the supported method first. More subscription capacity does not fix an unsupported rule.
Make one small test dictionary
Start with a few recurring terms and a short evaluation script. The current ElevenLabs app guide describes creating a dictionary, adding an input term, choosing Alias or Phoneme, entering the output and saving changes. It also warns that input capitalization matters. Use the documentation for the current interface rather than assuming every product surface has identical controls.
For API workflows, ElevenLabs documents uploading a PLS file and passing a dictionary/version locator to text-to-speech generation. The following is an illustrative alias-only PLS fragment for a team that has deliberately chosen to speak “CSV” as individual letters. It is a starting point to test, not a guarantee of a particular voice’s delivery.
<?xml version="1.0" encoding="UTF-8"?>
<lexicon version="1.0"
xmlns="http://www.w3.org/2005/01/pronunciation-lexicon"
alphabet="ipa" xml:lang="en-US">
<lexeme>
<grapheme>CSV</grapheme>
<alias>C S V</alias>
</lexeme>
</lexicon>Follow the programmatic dictionary guide for creation and attachment. Creating a dictionary and actually attaching the intended version to a generation are different steps. Keep both visible in your production notes.
Check the exact capitalization used in your approved scripts. If a term appears in several forms, handle the intended variants deliberately. Also listen for the rule affecting a longer word or an unrelated phrase. A global replacement that fixes one sentence but changes another is not ready for the shared library.
Treat numbers as meaning, not just pronunciation
A dictionary is particularly useful for recurring vocabulary. Numbers often need a different decision: what does this particular value mean? “2.4” could be a decimal measurement, a software version or a section reference. A single replacement rule cannot choose the correct meaning in every context.
ElevenLabs’ best-practice documentation discusses normalization of dates, amounts, phone numbers, units and other complex text. It notes that model behavior can vary and recommends making the intended spoken form clear. See the text-to-speech best practices.
For the software tutorial, the editor writes “version two point four” in the production script and “the sixth of July” when that is the approved date. An amount includes its currency. An identifier is read as an identifier rather than expanded into a large quantity. These are example editorial choices; select wording for the actual audience and meaning.
Keep the display copy and spoken script linked. A page can display a compact date while the narration uses an unambiguous long form. Review the two together so a wording correction does not accidentally change the offer, measurement or instruction.
ElevenLabs is a useful candidate when your content repeats specialized terms and the team wants a maintained script-to-audio process. A small dictionary and a representative listening test make that evaluation more informative than a generic voice demo.
Test terms inside complete sentences
Generate a compact test passage containing a term at the start, middle and end of an instruction. Include the capitalization and punctuation used in production. Listen for pronunciation, word boundaries, stress and the transition to the next word. A correct isolated name can still sound awkward in a full sentence.
Use the same voice, model and relevant settings intended for the published asset. Change one variable at a time. If the term improves after changing both the alias and the voice, the result does not tell you which change mattered. Keep the approved sample and a short reason for the chosen rule.
Ask a reviewer to listen without reading the script first. Can they identify the product name and understand the action? Then compare the audio against the text for omissions and altered numbers. The two passes answer different questions: intelligibility and fidelity.
For multilingual material, use a reviewer familiar with the target language. Reusing an English respelling across markets may produce the wrong local sound. Maintain language-specific decisions where needed rather than treating one dictionary as an automatic localization solution.
Version rules before revising published audio
A dictionary correction should have a reason and an affected-asset list. In the example company, revision 2 changes one product name rule. The editor lists the three tutorials that use that name, generates their affected sections and checks the transitions in each complete piece.
The ElevenLabs add-rules API returns a dictionary identifier and version identifier. Its reference also says that adding a rule for an existing matching input replaces that rule. Keep the returned version with the generation record so a future editor knows which pronunciation configuration produced the asset. See the dictionary rule API reference.
Do not assume editing a dictionary rewrites audio files that have already been generated or published. Treat those files as independent deliverables. Replace the approved asset deliberately, then play the final website or video output. Browser caching, a video export or an old CDN URL can otherwise leave the previous pronunciation in use.
Troubleshoot the narrowest cause first
- The rule has no effect: check the attached dictionary/version, exact input spelling and supported rule type for that model.
- The sound is closer but unclear: compare a shorter alias with the approved spoken form, then test it in context.
- A number is wrong: establish its meaning and rewrite the production script explicitly before adding a global rule.
- One sentence improves and another worsens: inspect the replacement scope and whether the same written term has two meanings.
- The preview is correct but the website is wrong: confirm the actual exported file and the page’s current audio URL.
Keep the last approved script, dictionary version and audio available until the revised asset passes review. If the correction creates a wider problem, restore the previous published asset and investigate the individual rule. That is a content rollback process, not a claim that a new generation will reproduce the old waveform exactly.
Approve the production asset as a complete package
Before publication, check pronunciation, factual meaning, captions or transcript, destination page and the rights required for the voice and generated output. Use the provider’s current permitted commercial workflow; a pronunciation feature does not change the licensing of a voice or source material.
A successful dictionary is a small, maintained collection of decisions the team can explain. It reduces repeated corrections and gives new editors a dependable starting point. Keep the final listening review: the goal is understandable, accurate narration for a real listener, not merely a saved rule in an account.
ElevenLabs and the ElevenLabs logo are registered trademarks or trademarks of ElevenLabs, Inc. Used with permission. Pagember is not sponsored by, endorsed by, or affiliated with ElevenLabs except as an independent participant in the ElevenLabs Creator Affiliate Program.
Program attribution requirements: Creator Affiliate Program terms.
Sources and verification
Editorial review 2026-10-19. Source check dates are recorded separately. Notes describe what each source supports and its limits.
View sources
- ElevenLabs pronunciation dictionary guideChecked Alias versus phoneme, current model compatibility matrix, app workflow and capitalization-sensitive input.
Limits: Model and surface support can change. Dictionary example was not executed. Pronunciation remains voice/context dependent; rights to source material, voices and commercial output must be established separately. - Using pronunciation dictionaries programmaticallyChecked PLS creation and generation dictionary/version locator; model-specific phoneme behavior.
Limits: Model and surface support can change. Dictionary example was not executed. Pronunciation remains voice/context dependent; rights to source material, voices and commercial output must be established separately. - Text-to-speech best practicesChecked Text normalization and context-dependent number/date/abbreviation meaning; output review.
Limits: Model and surface support can change. Dictionary example was not executed. Pronunciation remains voice/context dependent; rights to source material, voices and commercial output must be established separately. - Add pronunciation dictionary rulesChecked Existing matching input rule replacement; dictionary and version identifiers returned.
Limits: Model and surface support can change. Dictionary example was not executed. Pronunciation remains voice/context dependent; rights to source material, voices and commercial output must be established separately. - ElevenLabs Creator Affiliate Program termsChecked Required trademark/independence attribution, transcribed exactly from the existing project approved attribution file.
Limits: Model and surface support can change. Dictionary example was not executed. Pronunciation remains voice/context dependent; rights to source material, voices and commercial output must be established separately.