Skip to content
Hot storyPast story

Inworld uses an LLM to rewrite descriptions and iteratively filter synthetic voices

1 report1 sourceUpdated 12 d ago

What happened

AI overview

Inworld AI head of research Aleksey has published an experiment: Inworld Voice Design generates voices from text descriptions, then an LLM rewrites and combines those descriptions based on the evaluation results, iterating for 26 rounds. In the end, 214 voices passed the automated checks, and human reviewers picked 12 of them for its TTS voice library. Each round of generation measures audio quality and how much a voice differs from those already in the library. Inworld calls this approach "breeding" synthetic voices, with the aim of adding diversity to its TTS voice library; the full pipeline and audio samples were published on its Substack.

AI-written from the coverage · updated 12 d ago

Coverage timeline

Follow the coverage to see the story from different sides.

Nov 10
  1. Inworld AIPick
    Inworld breeds synthetic voices by having an LLM iteratively rewrite descriptions, picking 12 after 26 rounds

    Inworld head of research Aleksey used Inworld Voice Design to generate voices from text descriptions, then had an LLM rewrite and combine those descriptions based on evaluation results. After 26 rounds of iteration, 214 voices passed automated checks, and human review picked 12 of them for the voice library. Each round tests audio quality and how much a voice differs from those already in the library; the full pipeline and sample clips are published on his Substack.