ZONOS2 Voice Clone

This is the node that makes the pack worth installing. You give ZONOS2 Voice Clone a few seconds of someone talking and some text, and it synthesizes that text in their voice - audio-only zero-shot cloning, no reference transcript, no training run, no API call. Everything stays on your GPU.

It works the way IP-Adapter works for faces: the model doesn't look at the reference waveform at all. Instead a speaker encoder (an ECAPA-TDNN model) squeezes your reference clip into a single 2048-dimensional speaker embedding, and ZONOS2 conditions its generation on that fixed vector. Identity transfers, and because the embedding is a fixed point rather than a growing context, cloning is cheap and fast. The flip side, and the README is refreshingly honest about this: accent, cadence, emotion, and other utterance-level prosody don't reliably survive the trip. You get the voice, not necessarily the performance.

Wiring it up

The node takes three required inputs on top of the model:

• zonos2_model - from the ZONOS2 Model Loader.
• text - what the cloned voice says.
• reference_audio - a native ComfyUI AUDIO noodle. Feed it from a Load Audio node (the pack's example workflow uses LoadAudio). Best results need 5–30 seconds of clear, single-speaker speech; anything past 60 seconds gets clipped with a warning in the CLI, and very short clips get a recommendation warning too.

Two toggles genuinely change behavior:

• clean_speaker_background (default false) - the official clean/noisy recording flag. Leave it off for ordinary recordings; it's not a denoiser, it describes the source. Enable it only for genuinely studio-clean speech.
• accurate_mode (default true) - requests stricter adherence to the embedding, improving identity match at the cost of expressiveness. Turn it off for looser, more expressive delivery. It is not an accent-strength slider.

Below those sit the same sampling controls as the generation node (max_new_tokens, temperature, top_k, top_p, min_p, repetition settings, seed) and the same quality-conditioning buckets. temperature 0.8–1.0 plus a fixed seed is the "make it stop wiggling" combo.

The output is audio - mono, 44.1 kHz, native AUDIO - which wires into Save Audio/SaveAudioMP3 or downstream audio nodes. The speaker encoder is loaded lazily, so the first clone run will take a moment longer than plain generation.

Installing

Same pack as the other two nodes. ComfyUI Manager → ZONOS2 TTS, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Saganaki22/Zonos2_TTS-ComfyUI.git
../venv/bin/python Zonos2_TTS-ComfyUI/install.py

Restart. Needs Transformers 5.0–5.12 (4.x won't work), and the loader handles the model downloads. Also: read the model's acceptable-use policy once before you build anything. Cloning a voice you don't have permission to use is the one line this project draws hard.

Troubleshooting

• Clone sounds weak or wrong → it's usually the reference. Get 5–30 seconds of clean single-speaker speech with minimal silence, set clean_speaker_background to match the actual recording (not on by reflex), and keep accurate_mode on.
• Voice matches but the accent doesn't → model limitation, not a bug. The embedding is built for identity and discards prosody. Use a reference that clearly demonstrates the accent (ideally in the target language), compare a few fixed seeds, and know that accurate_mode and temperature can't conjure accent info the embedding never carried.
• CLI warning about a long reference → expected; it clips to the first 60 seconds.
• CUDA OOM → the FP8 preset from the loader, unload other models, lower max_new_tokens.

For getting one specific person's voice out of a short clip, this is currently the most direct path ComfyUI has - just manage your expectations on how they sound, not whose voice it is.