VoxCPM TTS

This is the only node in the pack you strictly need. Drop it, type some text, hit run, and you get speech - everything else in ComfyUI-VoxCPM (voice cloning config, advanced params, the training nodes) is optional extra plumbing that plugs into this one node. It's built on VoxCPM, a tokenizer-free TTS model out of OpenBMB/ModelBest: instead of chopping speech into discrete audio tokens like most TTS pipelines, it models speech directly in continuous space on top of a MiniCPM-4 language-model backbone. The practical upshot is expressive, context-aware prosody and genuinely good zero-shot voice cloning from a short sample.

The author, wildminder, isn't new to this - they also wrote the well-known ComfyUI-VibeVoice and ComfyUI-Chatterbox nodes, so this is someone who's been building TTS plumbing for ComfyUI for a while, not a one-off port. Community sentiment backs that up: people who've actually used VoxCPM tend to say it's underrated next to the louder names in the space (Chatterbox, Kokoro, VibeVoice) - more accurate, fewer skipped words on longer passages, and fast enough that the original model claims a real-time factor as low as 0.17 on a consumer RTX 4090.

Two models, four ways to use it

model_name picks the backbone: VoxCPM2 (2B params, 48kHz, 30+ languages, the one with voice design and reference cloning) or VoxCPM1.5 (800M params, 44.1kHz, mostly English/Chinese, the one with LoRA support). Default and the recommended pick is VoxCPM2.

From there this single node covers four modes depending on what you connect: plain zero-shot TTS with just text; voice design (VoxCPM2 only) where voice_design takes a plain-English description like "warm female voice with a slight British accent"; voice cloning via a connected voice_config carrying prompt or reference audio; and "ultimate cloning," which combines a reference clip (for identity) with a prompt clip (for prosody) for the closest match. Note that voice_design only does anything in plain-TTS mode - the moment reference or prompt audio is wired in, the model clones from that instead and ignores the description.

The inputs and outputs that matter

• text - what gets spoken. Multiline, so paste in a full script.
• cfg (default 2.0) - guidance scale. Push it up for tighter adherence to the voice description or reference; push it too far and it starts sounding less natural.
• steps (default 10) - diffusion steps. More can help quality, at the cost of generation time.
• normalize_text - text normalization for things like numbers and abbreviations. Off by default, and it needs the inflect and wetext Python packages installed to work at all.
• seed, lora_name (pick a trained LoRA from the dropdown, None by default), force_offload, device, dtype round out the rest.

Two optional inputs are where the modular part of the pack comes in: voice_config (from the VoxCPM Voice Cloning node) and advanced_params (from VoxCPM Advanced Parameters). Leave both unplugged and you get sane zero-shot defaults. The single output is Generated Audio, an AUDIO - wire it straight into a SaveAudio node or anything downstream that expects audio.

Installing it

Easiest path is ComfyUI Manager - search ComfyUI-VoxCPM, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/wildminder/ComfyUI-VoxCPM.git
cd ComfyUI-VoxCPM
pip install -r requirements.txt

The first time you actually run a node, it auto-downloads the selected model into ComfyUI/models/tts/VoxCPM/ - VoxCPM2 is a 2B-parameter model, so budget the disk space and expect a pause on that first generation.

Common issues & troubleshooting

device defaults to cpu in this node, not auto - if generation is crawling and you've got an NVIDIA card sitting idle, that's why. Switch it to cuda explicitly.

normalize_text errors out. It depends on the inflect and wetext packages, which aren't pulled in automatically - pip install inflect wetext into the same Python environment ComfyUI runs on fixes it. This is a genuinely common one; people hit it the moment they flip the toggle on.

First generation after a fresh install fails, second one works. A few users have reported exactly this pattern - if your very first run comes back empty or errors right after setup, just run it again before you start digging for a real bug.

VoxCPM1.5 output turns slightly metallic toward the end of longer sentences. This is a known artifact of the 1.5 model specifically, reported by people who've used it for longer-form narration. Switching to VoxCPM2 or breaking long text into shorter chunks both help.