How to build an ElevenLabs pronunciation dictionary

An ElevenLabs pronunciation dictionary gives speech generation reusable rules for words that need a particular pronunciation. In ElevenAgents↗ (opens in a new tab), open the agent, go to Voice Settings, add the dictionary, and save. An API workflow instead supplies a dictionary ID and version with the speech request. Test the exact voice and model before accepting a rule. fenbs can keep the spelling, approved reading, and test result so your next AI session starts with the same instructions.

How do you start a reusable dictionary?

1. List the names, abbreviations, and technical terms that need a particular reading. Preserve their capitalization.
2. For each term, choose an alias, a substitute spelling or phrase, or a phoneme rule, a pronunciation written with a phonetic alphabet.
3. For an API workflow, prepare the .pls dictionary described in the pronunciation guide and retain the dictionary ID and version returned by that workflow.
4. Attach the dictionary to the agent in Voice Settings, or pass its ID and version through pronunciation_dictionary_locators in the API request.
5. Generate a short test containing the exact term and listen before applying the rule to a larger piece of work.

The API pronunciation guide↗ (opens in a new tab) documents the dictionary file and request fields. The preparation and listening steps are a review method: they help you distinguish a rule you wrote from a pronunciation you have actually approved. Begin with one troublesome term so an error is easy to isolate.

When should you choose an alias or a phoneme?

An alias substitutes words or spelling intended to produce the desired reading. A phoneme rule describes the sounds using a phonetic alphabet. The distinction matters because phoneme support depends on the selected model; a dictionary being attached does not establish that every rule inside it will be applied.

Start with the simplest rule you can check by listening. For a hypothetical training script, an abbreviation might need to be spoken as individual letters rather than as a word. Record the intended reading in ordinary words before deciding which rule expresses it. This is an example of a review goal, not a guaranteed pronunciation for a particular voice.

Do not select a phonetic spelling simply because it looks precise. Have someone who knows the term confirm the result in a full sentence. A rule that sounds right alone can still be unsuitable for the passage you are recording.

Where do model and editor differences matter?

The current agent and API guides list several models with phoneme support, while the Studio documentation↗ (opens in a new tab) gives a narrower warning. Their descriptions of non-English phoneme support also differ. Check the guide for the surface you are using and preview your exact model instead of copying one compatibility list across every ElevenLabs tool.

The guides establish model- and product-specific support, rather than one universal compatibility list. If a rule does not produce the expected reading, check the selected model and product documentation before rewriting the entire script. Keep Studio editing and agent configuration separate when documenting your setup.

The Studio organization guide covers arranging the wider production. Dictionary rules solve a narrower job: giving a named term a checked reading. They do not replace decisions about speakers, chapter order, or the final script.

How should you test a pronunciation rule?

Use a short test script containing the term exactly as it appears in the final copy. Dictionary matching is case-sensitive, so a rule for one capitalization may not match another. Include a sentence before and after the term when the rhythm of the passage matters, and listen to the whole result.

• Record the exact source spelling, including capitalization.
• Keep the voice, model, dictionary ID, and dictionary version with the test.
• Compare the intended reading with what you hear, rather than approving the rule text alone.
• Test another occurrence from the real script before calling the rule reusable.
• Mark the result as accepted, rejected, or still waiting for a reviewer.

Change one thing at a time when a test fails: the rule, the spelling, or the selected model. Otherwise you can hear an improvement without knowing what produced it. Studio also documents speech generation as non-deterministic, so do not promise identical audio just because you kept the same recipe.

Keep the reviewer’s accepted reading in the shared record of the work so the next person knows which pronunciation passed, which version supplied it, and what still needs checking.

What should you keep after the rule is approved?

Save the dictionary source alongside a short approval record. Include the term, intended reading, tested voice and model, version, and a link to the approved take wherever you store audio. Retaining only the final recording makes later revisions harder because the reason behind a particular pronunciation is missing.

For slide narration, use the narration workflow to prepare the spoken copy, then apply the checked terms during speech production. A prompt library can also keep your reusable listening checklist. Neither substitutes for hearing the dictionary applied to the current script.

How to keep your checked pronunciation rules with fenbs

1. Create a free account.
2. Save each tricky word, how it should sound, and the rule that passed your listening check.
3. Add the tested voice, model, dictionary version, and a link to the accepted recording. Say what still needs a person to listen.
4. Connect ElevenAgents to fenbs if you are using its supported outside-tool connection. Studio does not have that same documented connection.
5. Save your next step before stopping so you can carry on from exactly where you left off, on any device and in any other AI.

ElevenAgents documents its custom MCP connection↗ (opens in a new tab), a way for an assistant to use outside tools. Workspace opt-in and feature restrictions apply. fenbs keeps the checked instructions and work record; it does not generate speech or install dictionary changes for you. If you only need one correction and no later handoff, keeping it in your current audio workflow may be enough. Explore more production workflows in AI at work.