VoxCPM Controllable Cloning

This is the middle child of the VoxCPM cloning family, and it's the one you'll actually use most. You drop in a short recording of a speaker, and the node reproduces their timbre - then you push the delivery around with a plain-English style instruction. Same voice, different performance. That combination is what "controllable cloning" means here, and it covers most real jobs: you have a sample of someone but no transcript, and you want them to say something new in a specific mood.

The reference audio does the hard work. The pack feeds it to VoxCPM2 as a timbre reference (the upstream code calls it "structurally isolated via ref_audio tokens"), and style_instruction rides along as a parenthetical control prefix on the text - the same bracket-trick Voice Design uses, but stacked on top of a real voice instead of a description. That's the whole division of labor: reference = who, style = how.

Inputs that matter

• reference_audio - the voice you're cloning. Quality here decides everything. Single speaker, clean speech, low background noise, no heavy music, no overlapping voices. The README is strict about this for good reason - a short clean clip beats a long noisy one every time.
• text - what the speaker says.
• style_instruction - the steering wheel. Good ones: slightly faster, cheerful tone, calm and steady narration, emotional, softer ending, slower pace, or 粤语口吻，轻松一点，像朋友聊天. Leave it empty and you get neutral-ish cloning.
• cfg_value (0–10, default 2) - how strongly the output follows the prompt/instruction. Push it up when the style isn't coming through.
• inference_timesteps, max_len, normalize, denoise - the pack's shared generation set. max_len up if audio cuts off; denoise for cleaning the reference before generation.

Output is a single audio socket for preview or save.

When to pick this over the alternatives

• No transcript, want a real voice + style control → this node.
• No transcript, want style control less and similarity more → Controllable Cloning still works, but clean up the reference.
• You have the exact transcript of the reference → step up to Ultimate Cloning. It gets higher similarity because it doesn't have to guess what the clip says.
• You don't want anyone's real voice at all → Voice Design.

Setup and troubleshooting

Shared pack install: ComfyUI Manager (search "ComfyUI-VoxCPM") or git clone https://github.com/starsFriday/ComfyUI-VoxCPM.git into custom_nodes, pip install -r requirements.txt, and download openbmb/VoxCPM2 into models/VoxCPM2. Heavy deps, Python 3.10–3.13, CUDA ≥ 12.

The failure modes are predictable. Result sounds off or lifeless → swap in a cleaner, shorter reference; that fixes more problems than any knob. Reference is noisy → enable denoise here and load the model with load_denoiser enabled on the loader - the denoise flag on this node does nothing if the denoiser was never loaded. And don't expect an exact impersonation from a five-second, music-backed clip; the model can only clone what the reference actually contains.