ZONOS2 Voice Generation (Emotion)

Sometimes you don't want to clone anyone. You just need a clean, natural voice to narrate a video or read some text aloud, and you want it to not sound like a GPS. That's the job of ZONOS2 Voice Generation (Emotion) - the simplest node in this pack and, because of that, the one worth trying first.

Here's the thing about this fork: the emotion controls only live on the Voice Clone node. Voice Generation has no reference audio, so there's no speaker vector to shift - no emotion dropdown, no emotion_cfg_scale. That's a deliberate design choice, not a missing feature. What you do get is ZONOS2's byte-level UTF-8 text handling (no transcript needed) across its roughly 40 supported languages, plus the full conditioning and sampling surface.

What it does

The node takes zonos2_model from the Model Loader and text (multiline), and returns mono audio as native ComfyUI AUDIO at 44.1 kHz. Wire the output into a Save Audio node, or feed it forward into an audio-driven avatar like MultiTalk or InfiniteTalk if you're building a talking head - that's the classic local pipeline in the audio ecosystem.

The controls that matter

For a first run, ignore almost everything. max_new_tokens (1024) caps how much speech you get - raise it if the model stops early, since it can emit an end-of-audio token before the cap. temperature 1.15, top_k 106, and min_p 0.18 are ZONOS2's official defaults and they're genuinely good; drop the temperature toward 0.8–1.0 if you want more stability across runs. top_p defaults to 0 (disabled) because this model uses Min-P, not Top-P - don't feel obligated to switch it on. seed is the one you'll touch daily: any positive value makes a run repeatable for identical inputs, while 0 means "surprise me."

Then there are the seven conditioning dropdowns: speaking_rate, loudness_lufs, estimated_snr, maximum_pause, estimated_bandlimit_hz, leading_silence, and trailing_silence. Each is a set of buckets the model was trained on, and "default" leaves that axis unconditioned. Trailing silence already defaults to the 0.25–0.5 s bucket, which is the official recommendation. Most people never touch the rest, but speaking_rate is the sleeper: pick a lower bucket and ZONOS2 reads slower, the closest thing this node has to a pace control.

Installing

Install and gotchas are identical to the loader: same pack, so ComfyUI Manager → search "ZONOS2 TTS (Emotion)", or clone plus the install.py helper, then restart:

cd ComfyUI/custom_nodes
git clone https://github.com/shiwano/zonos2-tts-comfyui-emotion.git
../venv/bin/python zonos2-tts-comfyui-emotion/install.py

Same Transformers 5.0–5.12 requirement (the #1 install killer for this pack - 4.x environments won't run it), same first-run multi-gigabyte model download, same 44.1 kHz mono output. And yes, the node name sets up a small letdown: "Generation" makes people expect emotion, but this is the plain-vanilla branch. If that disappoints you, the Voice Clone node is where this pack's actual party trick lives.

Reach for this node when you want a voice that just works - clean, expressive, multilingual - without hunting for a reference clip. It's the boring half of the pack, and boring is a feature.