Audiobook software gives each main character a steady voice, Audible says
The Audible team's text-to-speech system reads a whole book first, then assigns main characters fixed voices and writes delivery notes for each line. The tests are the team's own, described in an Amazon Science post.
Before it speaks a single word, a new audiobook system reads the whole book and works out who is in it. It then gives each main character a voice that stays the same from beginning to end, according to the Audible team that built it.
Arent van Korlaar, an applied scientist with Audible, writes in a post on the Amazon Science blog dated 23 October 2026 that ordinary text-to-speech models, which turn written text into spoken audio, do not reliably change speaking style between characters, as a human narrator would. Listeners can struggle to tell narration from dialogue, or to know who is speaking.
The post says listeners hear one voice for 10 to 15 hours per book, and general-purpose models often voice a character differently in chapter 1 and chapter 10.
The system starts with a component called BookLLM. It uses existing large language models to read the entire book and build a cast list. These are AI programs that work with text. BookLLM ranks the characters by importance and writes a voice description for each, using details such as age, gender, personality and role.
Each main character then gets a profile from a curated voice catalogue. The system deliberately makes the most prominent characters sound as different from each other as possible. Minor and one-off characters are read in the default narrator's voice.
For each line of dialogue, BookLLM also writes a caption, such as "whispering nervously", which the post compares to stage directions for the speech model. Because a character's voice profile is fixed once for the whole book, the character sounds consistent, while the captions add expression from moment to moment.
The team trained the system on expressive speech labelled with descriptions of how it sounds. The post says most existing data is neutral in tone, so the team built a captioner that listens to speech and writes such descriptions.
The team tested BookLLM on 25 English and Spanish books in the mystery, sci fi and romance genres, with two human annotators marking the correct answers. It identified which character speaks where, including telling narration from dialogue, with 97% accuracy in English and 94% in Spanish.
In a MUSHRA (multiple stimuli with hidden reference and anchor) test, a standard method in which listeners rate audio quality from 0 to 100, dialogue differentiation improved by 3.7 points.
Across English and Spanish, the team saw statistically significant improvements in listening experience, meaning the difference was unlikely to be chance, for 8 of 10 voices. The other 2 did not show a significant improvement. All of these figures are the team's own.
Sources
- Better speech synthesis for audiobooks · amazon.science