Scenema Audio Voice Clone ⚡

Every once in a while you don't need to generate speech from text - you need to take an existing recording and make it sound like a different person, without retraining anything. That's exactly what this node is: a standalone wrapper around the SeedVC identity-transfer bundle, the same offline post-pass that Scenema Audio Generate uses internally, exposed on its own so you can run the conversion on any audio you already have.

What it does

source_audio in, identity_reference in, converted audio out. SeedVC is a voice-conversion model: it keeps the content, rhythm and prosody of the source while transferring the timbre and identity of the reference. It's the "speak as X" tool, distinct from the diffusion generation path - no text, no prompts, no model loader needed. Just two clips and the conversion knobs.

The inputs

• source_audio - the recording whose content you keep. Whatever was said, however it was said.
• identity_reference - the clip whose voice you want. This is the fixed identity target.
• steps (25) - conversion sampling steps; more is slower and marginally more refined, the default is the sensible spot.
• cfg_rate (0.5) - how hard the conversion pulls toward the reference identity. Higher = stronger identity transfer, and too high starts to warp the source's natural delivery.
• seed - seeds the flow-matching noise so a given conversion is reproducible. Same inputs, same seed, same output.

All five are straightforward; for most uses the defaults plus a decent identity_reference are all you touch.

One output

audio (AUDIO) - standard comfy audio. It composes with everything, so you can chain it after any TTS node: generate with a generic voice, then transfer the identity you actually wanted. That's the workflow this node exists for - a lot of people run a fast, reliable TTS for the content, then use SeedVC for the voice.

How it relates to the big generate node

In Scenema Audio Generate, the SeedVC pass is built in (with identity_reference, skip_vc, vc_steps, vc_cfg_rate), and it's used to keep multi-chunk output voice-consistent. This node is that same conversion, standalone. The catch is the same in both places: the extras have to be on disk. SeedVC reuses Comfy's own Whisper and BigVGAN implementations plus a small checkpoint-specific architecture, so you need the seedvc, campplus, bigvgan, and whisper-small folders from the ChrisColeTech/scenema-audio extras repo under models/scenema-audio/extras.

Installing it

Part of the ComfyUI-GGUF-Loader pack under 🤖 CCTech/Scenema. ComfyUI Manager → search "ComfyUI-GGUF-Loader" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ChrisColeTech/ComfyUI-GGUF-Loader
cd ComfyUI-GGUF-Loader
pip install -r requirements.txt

Common issues

The two failure modes are both about the reference. First: a noisy or multi-speaker identity_reference makes the converted voice inherit that mess - keep the reference clean and single-speaker. Second: cfg_rate pushed too high (the range goes to 2.0) distorts the source's natural delivery; if the output sounds like the identity won but the emotion died, back it down toward 0.3–0.5. And if the node errors immediately, the extras aren't installed - that's the models/scenema-audio/extras folder missing, not the node. Same-shape, same-length, audible output is the contract here, so a failed conversion is loud about it rather than silent.