VoxCPM Ultimate Cloning

This is the node the pack saves its strongest trick for: continuation cloning. You hand it a clip of a speaker, type the exact transcript of that clip, and then type what they should say next - and the output should sound like the same person just kept talking. No cold start, no "new sentence" gap; the model continues from the prompt audio the way it continues from a prompt in any other generative model. It's the closest thing in this pack to "the speaker was in the room."

That quality comes with a price: you must transcribe the prompt audio accurately. This is not a convenience node. If you don't have the transcript and can't be bothered to write it, Controllable Cloning exists precisely for you.

How it works

The mechanism is in the names. prompt_audio is the spoken prefix, prompt_text is its transcript, and text is the continuation. The pack builds a prompt cache from the audio-plus-transcript pair (the upstream calls this the continuation/context mode - it's how VoxCPM2 does context-aware synthesis), then generates your text as if the speaker were continuing. Because the model knows what was said and how it was said, similarity is higher than timbre-only cloning.

• prompt_audio (required) - the clip the speaker "continues" from.
• prompt_text - the exact transcript. This is the one input you cannot fudge. The README is emphatic: wrong transcript quality directly hurts cloning quality, and it can degrade a lot.
• text - the new content.
• reference_audio (optional) - connect the same speaker again as an extra timbre reference to reinforce similarity. Worth doing if you have a second clean clip of them.
• Then the usual generation set: cfg_value (0–10, default 2), inference_timesteps (1–100, default 10), max_len (raise if output is too short), normalize, denoise.

Output is one audio socket.

The example that makes it click

prompt_audio says: 大家好，今天我们来聊一下配音技巧。
prompt_text:        大家好，今天我们来聊一下配音技巧。
text:               下面我会用三个例子说明停连和重音。

The generated audio should sound like the same narrator pausing and continuing. When it works it's spooky; when the transcript is wrong it's a completely different, garbled voice - there's no middle ground.

When to use it

• You need maximum similarity and you have the transcript → this is the node.
• The transcript is a pain to produce → use Controllable Cloning and accept slightly lower fidelity. For many projects that trade is correct.
• Reference is noisy → enable denoise and load the model with load_denoiser on the loader, or the denoise flag silently does nothing.

Install is the pack standard: ComfyUI Manager (search "ComfyUI-VoxCPM") or git clone https://github.com/starsFriday/ComfyUI-VoxCPM.git into custom_nodes, pip install -r requirements.txt, model files from openbmb/VoxCPM2 into models/VoxCPM2. And one etiquette note that applies to any cloning node: this stuff is exactly why voice cloning has a bad reputation - clone your own voice, or voices you have permission to use, and don't feed it nonconsensual samples.