Your China-based production partner has delivered a product film with the brand name pronounced correctly. A regional campaign now needs a different AI voice or model. Can the team carry the pronunciation dictionary into the next batch and skip listening approval?
ONCE recommends a small acceptance step before that batch starts. An approved recording proves what was accepted in that recording. To reuse the production setup, the team also needs to establish whether the new configuration produces the intended reading.
A dictionary rule and a working pronunciation are different deliverables
ElevenLabs published its pronunciation dictionary guide for Eleven v4 on October 1, 2026 and updated it on October 15. It distinguishes alias rules, which substitute text to guide speech, from phoneme rules, which specify pronunciation phonetically. It also separates dictionary phoneme rules from inline SSML: its guide says v4 supports dictionary phonemes and inline IPA, while it does not support SSML.
The official API guide warns that models without support for the relevant dictionary phoneme rules skip them and use default pronunciation. This is a specific vendor capability, not evidence that every voice platform behaves identically. The procurement implication is ONCE’s analysis: accepting a file and applying its intended reading need separate checks.
Buy a reusable pronunciation package with three connected records
A global marketing team and its manufacturing or production partners may agree on the product name while using different tools to render it. Keep the brand decision separate from the instructions used by a particular tool.
| Record | What belongs in it | Approval owner |
|---|---|---|
| Approved spelling | Names used on packaging, the website, captions and approved scripts, including capitalization variants actually used | Brand or product owner |
| Pronunciation instructions | Target-market reading, reference audio, tool-specific rules, language and configuration notes | Language reviewer and generation operator |
| Listening evidence | Rendered key lines, their scripts, voice and model identifiers, dictionary version and review date | Person authorized to approve that market’s audio |
This is a proposed ONCE handoff method, not a platform-mandated package. Where a service exposes no precise model version, record the available identifier, settings and generation date. Do not invent a version number to make the record look complete.
Keep pronunciation respellings out of the approved spelling record. Otherwise an editor copying generation input into captions may publish a production workaround as the product’s name. The pronunciation instruction can change while the customer-facing spelling stays approved.
There is also a concrete reason to inspect the rules rather than merely count them. ElevenLabs’ Best practices documentation says dictionary matching is case-sensitive and uses the first matching replacement in dictionary order. Check the forms that occur in the actual script and review overlapping entries.
Run a six-sample comparison before commissioning the batch
Consider a hypothetical overseas brand buying demonstration videos from a China-based supplier. Configuration A is approved; configuration B is proposed for the next campaign. This is an illustrative procedure, not a completed ONCE test.
Select three real lines from the upcoming script: an opening with the brand name, an instruction containing a model designation and acronym, and a closing action line that repeats the key name. Render all three under each configuration. The resulting six clips provide a compact comparison, with a record of how each was generated.
Ask the market reviewer to assess the names and abbreviations first, then sentence-level delivery and timing against the picture. An isolated word does not answer whether the full instruction still sounds natural or fits the product action. These six samples are a starting set; they do not certify every sentence or language.
Let the failure determine the repair
- If one capitalization fails, inspect coverage of the exact script form.
- If a phoneme rule is ignored, verify support for the specific model and interface before selecting another method.
- If pronunciation passes but the line no longer fits the shot, review pacing, wording or the edit.
- If a substitution changes an unrelated word, narrow the rule and revisit affected lines.
Generate the production batch after approval. If it fails, an existing approved recording or a human pickup may be an option, subject to script, rights and synchronization checks. The repair should address the identified problem rather than automatically rebuilding every language version.
Make reapproval part of the ongoing service
For a single film, the purchase can specify the approved mix and clean voice track. For an ongoing content service, also name the custodian of the dictionary, the person approving new terms and the person triggering a review after a voice, model or interface changes.
Separate maintenance, generation, language review and replacement in the finished edit when defining scope. There is no universal fee implied here. The useful question is whether the quoted service includes the work needed to keep later deliveries consistent, and who authorizes a change.
For ONCE’s progression toward managed content services, the asset is the organized set of decisions and references that a later production can use. Reuse should reduce repeated briefing; it should not turn an old approval into a guarantee about future output.
Before the next order, send your partner approved names, existing reference recordings, target-market requirements and the proposed generation configuration. Request the key-line comparison before authorizing bulk production. Broader planning for language versions can start with ONCE’s video production articles; this acceptance step addresses whether the pronunciation setup is ready for reuse.
Source scope
The vendor guide above is dated October 1, 2026, updated October 15. The linked API and best-practices documentation did not provide a precise publication date in the material reviewed; all were checked on November 2, 2026. Product surfaces can have different support scopes, so confirm the actual interface and model. No audio was generated for this article, and the proposed acceptance method is not an industry adoption claim.
