😻 Kitten-TTS

The name is pure vibes, but KittenTTS is real: this single node drops a sub-25MB, 15-million-parameter text-to-speech model into your ComfyUI graph and turns a text string into spoken audio on CPU. No GPU, no API key, no cloud. You type a line, pick one of eight voices, and out comes a raw 24kHz waveform. For a diffusion rig where audio is usually an afterthought bolted on via lipsync pipelines, that's genuinely handy - you can generate narration or dialogue for a video workflow without leaving the graph.

The honest caveat up front: this is a developer preview model. The HF card says so itself, and the Reddit reaction to KittenTTS has been "wow, that size" followed by "…but it sounds like it's speaking through paper." It's remarkably good for 24MB. It is not ElevenLabs. If you need broadcast-quality VO, look elsewhere; if you need a quick voice track on a local machine with no internet once the model is downloaded, this is one of the few options that just works.

How it works

Under the hood this is a wrapper around KittenML/kitten-tts-nano-0.1, an ONNX export of the KittenML nano model. The pipeline is: your text gets phonemized to IPA via espeak-ng, tokenized into the model's vocab, then pushed through an ONNX session along with a style embedding pulled from voices.npz (each of the eight voices is just a vector in that file) and a fixed speed of 1.0. Out the other end comes a numpy waveform. That's the whole thing - one small ONNX file plus a few KB of voice embeddings, no big PyTorch stack. The CPU-only design is the whole point: it's built to run "literally everywhere," per the model card.

One thing worth knowing: the node returns a bare numpy array, not a saved file and not a fancy audio object. It's labeled AUDIO so it will try to wire into whatever audio nodes your ComfyUI build has, but there's no built-in preview or save button here - plan to grab the waveform and write it out with your own audio node, or script it directly with the model's generate_to_file.

The inputs that matter

There are only two, and you'll set both every time:

• text - the string to speak. Plain English is the safe zone; the phonemizer is an espeak-ng English-IPA step, so heavy non-English text will get mangled.
• voice - pick from expr-voice-2-m/f through expr-voice-5-m/f. The m/f is the gendered register, and 2→5 are the different character timbres. There's no clone-your-own option yet; you get these eight and you like it.

Installing it

ComfyUI Manager is the easy path: search "KittenTTS" and install. Or clone it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/Lovzu/ComfyUI-KittenTTS

then restart ComfyUI and let Manager (or the pack's requirements) resolve dependencies. Fair warning: the requirements.txt is not lightweight - spacy, onnxruntime, misaki, soundfile, num2words, huggingface_hub, and espeakng_loader, which is what provides the espeak-ng engine and its Python module. If you're already in the custom-node dependency-hell spiral, this pack adds one more layer of it. The model itself (~24MB) downloads automatically from Hugging Face on your first run, so you need internet exactly once; after that it's fully local.

Where people get burned

Two things, and both are real as of the current code.

First, the packaging bug: the shipped generate_voice.py defines a function misspelled genarate_audio, while nodes.py imports generate_audio. The node loads fine, then the first run dies with ImportError: cannot import name 'generate_audio'. The one-line fix is in the module file - rename the function to generate_audio (or change the import to match the typo, your call). If you hit this, it's not you; it's a typo.

Second, first-run dependency failures. If your first execution dies with No module named 'espeakng' or a missing-binary error, the espeakng_loader install didn't land properly - reinstall the requirements and restart ComfyUI before assuming the node is broken.

And third, lower your quality expectations before you run it. It's a 15M-parameter preview model. It'll read your line clearly, and then you'll notice the buzzy, papery texture on anything long. For a test track or a placeholder voice it's perfect. For anything you'd put in front of a client, budget time to swap in a bigger TTS.