VibeVoice

Microsoft's research text-to-speech model, run locally, cloning a voice from a sample.

VibeVoice is Microsoft's open-source research TTS model. It clones a voice from a reference sample and streams audio as it is generated.

Experimental, admin-only and a research preview: heavy on the GPU, and some lines come out with a second or more of babble before the actual speech.

Setup

Add the service

Manage Modules → Add Modules → VibeVoice → Install. Voxta installs the Python runtime, FFmpeg, the model and the voice sample library (watch the Terminal for progress).

Pick a voice

Choose from the voice library, or give a character its own sample. Set Default Female / Male Voice in the module settings for characters without one.

Models

Only the transformers conversions (-hf) load.

Model	VRAM (weights)
VibeVoice-1.5B (default)	~5 GB
VibeVoice-7B	~17.4 GB

Allow about 0.6 GB more at runtime. Use Cuda off runs on the CPU.

Settings

Setting	Notes
CFG Scale	1–2, default 1.3. Higher sticks closer to the reference voice; too high sounds stiff.
Diffusion Steps (advanced)	2–30, default 10. Lowering it saves little time for a noticeable loss in quality.
Thinking Speech (advanced)	Short sounds played while the reply is being generated.

Voice samples

The sample is the whole acoustic condition: VibeVoice copies the room and background it was recorded in, not just the voice. Use clean, dry recordings.
Only the first 20 seconds of a sample are used.
Supports affect clips recorded beside the voice file.