🎤 VibeVoice Preset Maker (0.5B)

Let me save you a few hours: this is the one node in the whole AIIA pack whose own README tells you not to use it. It's marked experimental, and the warning isn't a CYA disclaimer - it's an accurate description. If you found this page because you want to clone a voice with the 0.5B realtime model, step back now and read the "what to do instead" section at the bottom.

VibeVoice Preset Maker (0.5B) is supposed to take a reference audio clip and bake it into a .pt voice preset for the VibeVoice-Realtime-0.5B model - the same kind of preset file Microsoft ships for its official voices (Carter, Emma, the multilingual ones). In theory that's your path to custom voices on the fast, low-latency streaming model.

How it works (and why it doesn't)

It validates that you've loaded the 0.5B realtime model (it refuses to run with 1.5B/7B), processes your reference audio through the model's streaming audio processor, and writes out a preset file. Inputs: vibevoice_model (from the AIIA VibeVoice Loader, must be the 0.5B realtime one), reference_audio, and preset_name (default my_new_voice_preset). Output: preset_path, a VOICE_PRESET you'd wire into the realtime TTS node.

The failure mode is the model, not the node. Community testing and the author's own notes agree: the 0.5B weights appear to have been trained (or regularized) around the official preset voices and don't generalize to arbitrary custom timbres. Even with a full minute of clean, high-quality reference audio, generations frequently loop, degenerate into gibberish, or come out as noise. It's not something you can tune around with a longer sample.

What to do instead

Three honest options, in order:

1. Use the official Microsoft presets. Download en-Carter_man.pt, en-Emma_woman.pt, the JP/KR ones, etc. from the Microsoft VibeVoice repo into models/vibevoice/voices/streaming_model/ and use them with the realtime node. That's the supported path and it works.
2. Clone with the Standard node instead. The AIIA_VibeVoice_TTS (Standard) node, paired with the 1.5B or 7B model, does proper zero-shot voice cloning from a reference clip - no preset step, no instability. This is the path for actual custom voices.
3. If you're a researcher poking at the 0.5B's zero-shot limits, sure, this node is your lab bench. Just know you're debugging the model, not your workflow.

Install

It's part of the pack, so installation is the pack install plus the 0.5B model:

cd ComfyUI/custom_nodes
git clone https://github.com/havvk/ComfyUI_AIIA.git

Then models/vibevoice/microsoft/VibeVoice-Realtime-0.5B/ needs the model files and the Qwen2.5-0.5B tokenizer files (tokenizer.json, tokenizer_config.json, vocab.json, merges.txt) - the README has the exact wget commands, because the HF download sometimes skips them.

Common issues

The only real "issue" is the one you've already read. Don't chase it. If a tutorial online told you to clone a voice with this node, it was written by someone who hadn't run it.