Skip to content

VoiceCraft

Edit speech and clone voices with a neural codec model

Visit official site

What VoiceCraft is

The model is described as a neural codec language model based on token infilling for speech editing and zero-shot text to speech. With only a few seconds of audio from an unseen voice, it can clone that voice or edit a recording.

What you can do with it

  • Edit speech content
  • Perform zero-shot text to speech
  • Clone voices the model has not seen before
  • Edit existing recordings

Application scenarios

  • Generate natural speech for an audiobook or a podcast episode
  • Edit a recording to change content or the speaker voice
  • Clone a voice from a small audio sample for custom speech content

Model notes

  • The approach is based on a neural codec language model with token infilling
  • Audiobooks, online videos and podcasts are named as in-the-wild data
  • Only a few seconds of audio are needed for an unseen voice

Pros & cons

✓ What we like

  • An unseen voice can be cloned from a few seconds of audio
  • Both speech editing and text to speech are covered

! What to watch out for

  • The model is described for in-the-wild data such as audiobooks, online videos and podcasts
  • Editing and cloning start from a supplied audio sample

FAQ

How much audio is needed for a new voice?

The material states only a few seconds of a speech sample are needed for an unseen voice.

What data is it described for?

Audiobooks, online videos and podcasts are named as in-the-wild data.

Last reviewed: 2026-11-15

More AI audio tools tools

View all →

How we review