A model can pronounce a name correctly in one sentence and switch pronunciation later. Test the same unfamiliar word at a beginning, within a sentence, and after a list. Consistency matters when listeners must associate the name with a character, feature, or speaker across a full episode.
Build the test input
Invent a character called โRavenaโ and select one approved pronunciation. Put it in a greeting, a longer explanatory line, and a sign-off. Write a short approved audio reference or phonetic note. Avoid real protected names so revisions can focus on clarity and length rather than brand policy.
Run the proposed comparison
In the baseline, use the same original spelling without extra cues. In a separate condition, use each operation's documented speech guidance if available. Review the baseline and corrected outputs without revealing which version was made with pronunciation help. Collect both exactness and perceived naturalness ratings.
Decide what passes
Count the number of occurrences that match the approved reading before any manual edits. A model that fails in only one sentence can pass on a small clip after replacement but may be costly at scale. Flag a correction spelling that changes a nearby word or produces a distinct accent.
Act on the outcome
Use a reviewed pronunciation list when the production covers multiple episodes or localization languages. Cross-check important names after each script edit. Where synthesized pronunciation remains unstable, record a human pickup or rearrange the sentence before selecting the model for the whole script.

