Clone and locally edit speech with emotion controls

ViiTorVoice is a speech generation workflow for voice cloning, local speech editing, and expressive delivery controls. You can provide a reference voice, generate new speech from text, or take an existing recording and replace only the part that changed.

That local edit workflow is the clearest reason to pay attention. For example, if a voiceover says "send the notes before Friday" and you need it to say "before Monday," the workflow is designed to find that changed region and regenerate the replacement span instead of forcing you to recreate the whole clip from scratch.

It also supports emotion and non-verbal tags, so you can test whether a line should sound happier, sadder, breathier, or more conversational. Treat those controls as production checks, not automatic polish: the real test is whether the edited line still sounds like the same speaker.

Start with the Hugging Face Space if it is running, then test a short edit rather than a long script. Use one clean reference clip, write the original transcript, change one phrase, and listen for whether the edit blends into the surrounding audio.

For local reproduction, use the public GitHub project and download the ViiTorVoice-NAR weights into the expected `local_models/` folder. The repo exposes a local HTTP service for voice cloning and text-local-edit requests, so the workflow is repeatable without relying only on the hosted Space. The setup is still a speech generation stack, not a one-command browser toy, so use the demo as the first filter before installing it locally.

The public Space is a client for a backend service, so exact hosted behavior depends on that service being available. The reproducible path is the open model weights plus the GitHub service setup, and local results should be checked carefully for edit seams, pacing shifts, and unnatural emotion control.

• Patch a short voiceover line without rerecording the full take.
• Test cloned voice drafts for narration, product demos, or internal prototypes.
• Compare neutral delivery against emotion-tagged versions of the same sentence.
• Prototype localized or corrected speech clips from one reference voice.
• Build a local speech editing workflow when repeatability matters.

Try the demo

View model page