The system uses natural-language captions such as “whispering nervously” to direct delivery, and creates a book-level voice profile for each character so that voice can remain consistent across chapters.
Amazon says its character-profiling and dialogue-attribution component identifies dialogue with 97% accuracy in English and 94% in Spanish. In listener tests, rendering distinct character voices increased dialogue differentiation by 3.7 points on the 0–100 MUSHRA scale. The company reports statistically significant listening-experience gains for eight of ten tested voices across English and Spanish.
The work combines newly recorded and human-captioned data with synthetic data, including captions derived with expert models and signal-processing methods. Amazon describes research results rather than a named public product. The source does not establish public availability, an API, pricing, model versions, code or dataset release, or whether the system is deployed in Audible audiobooks.