VoiceCraft is an open-source speech model for zero-shot speech editing and text-to-speech, developed by researchers and published on GitHub, where it has gathered around 8,600 stars. Instead of treating audio editing and voice cloning as separate problems, VoiceCraft frames both as a single autoregressive generation task over neural audio codec tokens, which lets it produce remarkably natural results from very little reference material.
The model works by masking out a span of an existing recording and regenerating it from the surrounding context and a target transcript, so words or phrases can be replaced, inserted, or deleted while the speaker’s voice, prosody, and recording environment stay consistent. The same architecture doubles as a zero-shot text-to-speech system: give it a few seconds of reference audio plus text, and it synthesizes new speech in that voice. VoiceCraft was trained on large-scale real-world speech data, which makes it robust to the noise, accents, and informal speaking styles found in podcasts, interviews, and user-generated content rather than only clean studio recordings. The repository ships inference code, pretrained checkpoints, and notebooks so researchers and developers can run both editing and TTS pipelines locally.
VoiceCraft is free and open source. The code and model weights are published in the GitHub repository for research and experimentation at no cost; you only need your own GPU compute to run it.
VoiceCraft suits speech researchers, podcast and audio producers who need surgical fixes to recordings, and developers prototyping voice-cloning or dialogue-editing features who want a strong open model rather than a paid API.