OmniVoice Generate

This is the node that actually makes sound. OmniVoice Generate Audio takes the pipe from OmniVoice Load Model and your target text, and runs it through k2-fsa's OmniVoice - the local, zero-shot multilingual TTS that's been giving cloud services a run for their money. Same idea as a KSampler: the load node staged everything, this one does the sampling. Wire its audio output into PreviewAudio or a VHS save node and you're done.

How it works

OmniVoice is a diffusion language model, and the node picks your mode for you. If the pipe carries reference audio, you get voice cloning - your words, someone else's voice. If there's no reference, you get voice design: the four style dropdowns (style_gender, style_age, style_pitch, style_accent) get joined into a short instruction like "female, low pitch, british accent" and handed to the model. The gotcha is the fallback: leave every style on none and it defaults to a plain male voice, so don't expect a lottery - expect "male" until you tell it otherwise.

Everything else is diffusion sampling passed straight through: num_step (32 default, more steps = cleaner but slower), guidance_scale for classifier-free guidance, and deeper knobs like t_shift, layer_penalty_factor and the two temperature values that you can honestly ignore until you're chasing artifacts. seed makes runs reproducible, language can be left on auto and the model infers it from the text.

The inputs that matter

• text - what it says. That's the whole job.
• language - auto handles English, Chinese, Japanese, Korean, French, Spanish, German, Portuguese, Russian, Arabic, and Hindi. Set it explicitly only if auto guesses wrong.
• style_gender / style_age / style_pitch / style_accent - voice design only. "No ref audio only," per the author's own tooltips.
• num_step / guidance_scale - the two you'll actually touch. Bump steps if output sounds muddy.
• speed - 0.25× to 4×, handy for matching a video's pacing.
• use_duration / duration - flip the switch and the output is padded or trimmed to a target length. Off by default.
• postprocess_output - on by default; it trims silence and smooths the ends, so only disable it if you're chasing a specific artifact.

Outputs: audio (standard ComfyUI AUDIO, 24 kHz) and a status string that tells you the mode, seed, and sample count - nice for checking what actually happened in a big graph.

Installing it

Same story as the load node: ComfyUI Manager, search ComfyUI-OmniVoice_CRT, or clone into custom_nodes and run install_omnivoice_safe.bat on Windows (it installs omnivoice>=0.1.0 with --no-deps plus huggingface_hub). On Linux/Mac, pip install those by hand. First run downloads the OmniVoice model from Hugging Face into ComfyUI/models/omnivoice.

Where people get burned

Two hard errors right in the code: empty text raises, and a missing pipe raises - both with messages that tell you exactly what to fix. Real-world trouble is subtler. With short or muffled reference clips, OmniVoice tends to skip words or hallucinate, and community reports echo that; a clean few seconds of audio plus an accurate transcription override fixes most of it. The wrapper generates one utterance at a time - the README's "not implemented" list admits batch multi-utterance generation isn't there, so loop it or live with single shots. And expect a real VRAM footprint, roughly 5–6.5 GB while generating; offload_after_generate on the load node frees it afterwards. OmniVoice is fast for what it is - people measure it in the tens of times real-time on a 5090 - so the waiting is mostly the model load, not the speech.