Boyo Voice Enhancer

Want your generated character to sound like a specific person, or to clean up a voice track before it hits your video saver? Boyo Voice Enhancer runs Seed-VC - the zero-shot voice-conversion diffusion model from Plachtaa - entirely inside ComfyUI. Feed it the audio you want to change and a short reference clip of the target voice, and it resynthesizes the speech in that voice. No API key, no cloud, no per-minute pricing; it just needs a decent GPU and a patient first run.

This is the node that makes the pack's TTS story make sense. Boyonodes already generates speech with Chatterbox Turbo; this is the other half - converting or cleaning the result. In a dubbing or lip-sync workflow the shape is usually: TTS → (audio padder for timing) → Boyo Voice Enhancer → save audio → mux with video.

How it works

On first run it imports the Seed-VC wrapper, picks your device (CUDA, then MPS on Mac, then CPU as a sad fallback) and loads the models - this is the "may take a moment" step, and it downloads weights the first time. Both your video_voice (source) and target_voice (reference) AUDIO tensors are resampled to 22,050 Hz, written to temp WAVs, and handed to Seed-VC's convert_voice. What comes back is a new AUDIO tensor at that same 22,050 Hz, ready to wire onward.

A frank note on the knobs, because the node's face is busier than its internals: the parameters that actually reach Seed-VC are diffusion_steps (default 30 - more steps, cleaner conversion, slower), length_adjust (stretch or compress the output duration), and similarity_cfg, which maps to Seed-VC's inference CFG - crank it toward the reference voice and it clings harder to the target. intelligibility_cfg, convert_style, top_p, temperature, and repetition_penalty are accepted inputs in this build, but they're not threaded into the conversion call, so don't expect them to do much yet. Start with the three real ones.

Install

This is the high-maintenance node in the pack. The pack itself is on ComfyUI Manager ("Boyonodes") or:

cd ComfyUI/custom_nodes
git clone https://github.com/DragonDiffusionbyBoyo/Boyonodes

But Seed-VC is a git submodule, and a plain clone leaves that folder empty - which is the #1 way this node fails. You need:

cd Boyonodes
git submodule update --init --recursive
pip install hydra-core omegaconf munch descript-audio-codec
pip install -r requirements.txt

That requirements file is the heavy one from the README - librosa, transformers, soundfile, and friends. The repo ships a Portable auto_install_nodes.bat if you're on Windows and want the whole dance scripted.

Where people get burned

The empty submodule is the classic - you'll get an import error the instant the node runs, not at startup, so it can look like the workflow is fine until the last second. First run also downloads models, which on a slow connection can look like a hang - let it sit. Output is fixed at 22,050 Hz, so if the rest of your pipeline expects 48k, resample after. And on a CPU-only box this will be glacial; it's a diffusion model, treat it with respect. Seed-VC is real and well-regarded, but it's not magic - short, clean reference clips give you dramatically better voice matching than long noisy ones.