An AI narrator can read a smooth paragraph and still mispronounce the one word that matters most. For a garden tour, that word may be a plant’s Latin genus, a donor’s surname or a street name visitors use to find an entrance. ElevenLabs pronunciation dictionaries provide reusable rules instead of requiring the editor to invent a different spelling in every script. This practical guide builds a small, reviewable dictionary for a fictional botanical tour, compares alias and phoneme rules, and sets out how to check exports. It does not claim that a real account or ElevenLabs voice model was tested for this article.
Separate verified terminology from synthesized speech
Start with the garden’s approved plant list and signage, not the AI output. Copy the exact botanical names from a reviewed inventory; decide whether the recording should speak a genus, a common name or both. Plant names may vary in pronunciation by local practice, so ask the garden’s specialist to choose the preferred audience-facing form before encoding a rule. Add one fictional donor surname as a safe illustration of why consistency matters, but never use a real donor’s voice or private data without permission. Build a pronunciation sheet with the written spelling, accepted spoken version, language, reviewer and a short example sentence. This source sheet is the editorial authority; the dictionary is merely a production mechanism.
Choose alias or phoneme rules for the actual model
ElevenLabs documents two major kinds of rules: aliases that substitute a pronunciation-friendly written string, and phoneme rules using IPA or CMU notation. Their compatibility varies by speech model. The October 2026 Eleven v4 guide describes both dictionary rules and a pronunciation panel in the Text to Speech experience, while the developer documentation specifies which models accept phoneme tags and when aliases are safer. Avoid asserting that any phoneme entry works universally. For a model without the relevant phoneme capability, use an alias where supported and listen carefully for side effects. An alias can make a technical word sound closer to the intended form, but it can also change how text is spoken if the substitute is ambiguous.
Build a compact glossary with exact capitalization
Create a small set of rules rather than importing a hundred unreviewed terms. The documented dictionary formats and workflows treat matching as case-sensitive, so “Salvia” and “salvia” may need separate entries depending on the application. Keep one accepted variant for each context and record why it was chosen. For example, a plant genus in a scientific segment may retain the Latin form, while a visitor directions segment may use an accessible common name instead. Never manufacture an IPA string simply because it looks scholarly; confirm it with a knowledgeable speaker or reference before adding it. Maintain a clear source for each pronunciation and mark unresolved names as “human review required.”
Preview each rule inside a natural sentence
A single isolated word can sound correct even when a narrator stresses it badly inside a paragraph. Use a short test phrase such as “The next bed contains Salvia beside the south path,” then listen for the word and the sentence rhythm. The current ElevenLabs guide describes previewing a rule with the selected voice; use that option when available. Retain a baseline clip and an updated clip in the same voice and settings so reviewers can notice whether surrounding words changed. Do not label those example clips as performance benchmarks unless a controlled test actually took place. The success condition is an approved pronunciation in context, with readable wording and no unintended change to the factual script.
Keep dictionary rules from altering quoted source material
Some narrations quote a scientist or volunteer. A pronunciation rule should not silently rewrite what a speaker said in a real recording, and a text-to-speech quote must still match the cited statement. Keep the original quote as written source text, and apply pronunciation only to the generated narration track. If the donor’s name has two legitimate spoken forms, ask the person or authorized representative which form the production should use. Record that authorization. Watch out for short aliases that collide with ordinary words elsewhere in a project. A glossary rule intended for one tour could make a separate exhibit sound wrong. Recheck repeated terms after a dictionary update because a global rule can affect more clips than expected.
Version the dictionary before the next generation pass
Treat the pronunciation dictionary as a versioned production asset. ElevenLabs provides an API workflow with dictionary IDs and version references, and a product interface for managing rules in supported experiences. Name the approved list after the tour, record the model and voice configuration, and save the version that was used for each rendered segment. A changed rule can require new audio exports even when the script has not changed. Do not assume changing a dictionary automatically repairs published files or external captions. Flag every affected clip, regenerate it, and recheck captions and sentence timings. A tiny version table is much more useful to a production team than an untraceable collection of edited phonetic spellings.
Proof the entire tour on visitor equipment
After the terminology reviewer signs off, listen from beginning to end on the type of device visitors actually use. Check whether plant labels appear in captions with the correct spelling even when a phonetic alias was used for audio. If a botanical term is also a link in the digital guide, ensure the link text remains searchable and accurately matches signage. Confirm that the spoken route directions are operational, not only linguistically pleasing. A correct pronunciation of a nonexistent gate is still bad guidance. The educational diagrams here model the approval sequence; they are neither screenshots of ElevenLabs nor the results of a real speech-quality comparison.
Publish a reusable and reversible handoff
Hand over the approved vocabulary sheet, dictionary version, plain script, audio exports, subtitle files, language and voice settings, and reviewer sign-off. Document which terms were intentionally left in their default pronunciation. If a name is corrected later, update both the source glossary and all affected output clips, then archive the superseded files so team members do not accidentally reupload the wrong recording. Be especially careful with personal names and language communities: preferred pronunciation is a matter for the people represented, not solely an algorithm. The valuable output is a stable pronunciation policy that new editors can repeat without guessing and without claiming independent verification of every possible TTS model.
How We Research
This ElevenLabs Pronunciation Dictionary guide is independent editorial guidance grounded in official documentation: ElevenLabs: pronunciation dictionary guide, October 2026; ElevenLabs: pronunciation dictionaries documentation; ElevenLabs: TTS best practices. The illustrated scenario is a fictional narrated botanic garden tour with labels for Salvia, Echinacea, two garden locations and a deliberately unusual fictional donor surname. We checked the vendor-described capabilities and important eligibility conditions, then mapped a sequence of editorial reviews that readers can repeat with their own authorized media. The images are original teaching diagrams, not product UI screenshots, generated exports, results of a live account test, or measured benchmarks. We did not upload private recordings, access a paid vendor account, run a real A/B experiment, measure performance, or verify hypothetical timecodes. Retain source assets, check current plan availability and privacy settings, and review exported media and metadata before publication.
Frequently Asked Questions
An alias substitutes text; a phoneme rule supplies a pronunciation notation such as IPA or CMU where supported.
No. Compatibility varies; check the documentation for the exact model and rule type.
Dictionary matching can be case-sensitive, so separate written variants may require rules.
The notation should be verified with a knowledgeable speaker or pronunciation reference first.
No. Captions generally should retain the approved written name, while the alias controls only the spoken output.