Skip to content
Trending storyDeveloping

Audible uses controllable TTS and BookLLM to give audiobook characters distinct voices

2 articles2 sourcessince Oct 23Last article 2h ago ·

Overview

AISummary of 2 articles

Audible's technology team uses controllable text-to-speech with its BookLLM component to voice audiobook characters with distinct, consistent voices, according to Amazon Science.

BookLLM reads the entire text first, builds a cast list, ranks characters by importance and derives a voice description for each. Minor and one-off characters fall back to the default narrator.

Amazon Science reports that BookLLM identifies speakers with 97% accuracy in English and 94% in Spanish. It says the approach improved dialogue differentiation by 3.7 points on a MUSHRA test, and that listeners preferred the new audio for eight of ten voices in statistically significant tests across English and Spanish.

A later Amazon Science post on X repeats the 97% English accuracy figure, describing the system as profiling characters and attributing dialogue to the right speaker.

Written by AI from the articles below · updated Nov 2, 10:47 PM ET

Check the sources:

Latest developmentNov 2, 10:21 PM ET
Audible's controllable TTS attributes dialogue to characters with 97% accuracy

Article timeline

The articles in this story. Times are ET.

Nov 2
  1. Amazon ScienceOfficial
    Audible's controllable TTS attributes dialogue to characters with 97% accuracy

    AIAmazon Science says Audible's controllable text-to-speech system profiles characters and attributes dialogue to the right speaker with 97% accuracy in English. The system then voices each character accordingly. The post links to a blog post with further details.

Oct 23
  1. Amazon ScienceOfficial
    Amazon's Audible team uses controllable TTS to voice audiobook characters

    AIAudible's technology team is using controllable text-to-speech to give audiobook characters distinct voices, with its BookLLM component identifying speakers at 97% accuracy in English and 94% in Spanish. The approach improved dialogue differentiation by 3.7 points on a MUSHRA test, and listeners preferred the new audio for eight of ten voices in statistically significant tests across English and Spanish.

Heat Trend

Not enough continuous observations to show a trend yet.