OmniVoice TTS

OmniVoice is the open TTS that people actually got excited about in 2026 - a diffusion-based voice model from k2-fsa that does zero-shot cloning in 600+ languages, on local hardware, and fast enough that the usual complaint is it's "lightning quick compared with VibeVoice" rather than slow. This node is the other half of the omnivoice_comfy pack: it takes the model object the OmniVoice Loader gives you and turns text into an audio clip, with or without a reference voice.

Wire it up like this: OmniVoice Loader → OmniVoice TTS → any audio output node (PreviewAudio, SaveAudio, a VHS video node if you're dubbing a clip). The output is a standard ComfyUI AUDIO object, so it feeds everything else normally.

The inputs that actually matter

The node has a long input list, but most of it is pre-tuned defaults you should leave alone. The ones you'll touch:

• text - the words to speak. This is the target sentence, full stop.
• language - defaults to English. OmniVoice genuinely covers hundreds of languages, so if you're going multilingual this is where you say so.
• instruct - optional style direction ("male, australian accent, low pitch" per the placeholder). It shapes delivery and tone, not the words.
• ref_audio + ref_text - the voice cloning pair. Feed a reference clip into ref_audio and put its transcript into ref_text. Both or neither: hand it audio without text and it throws Voice cloning requires ref_text - Whisper auto-transcription is deliberately disabled in this pack.

The rest - num_step (28), cfg (3.0), t_shift (1.0), denoise (off) - are the author's tuned baselines. They're exposed because someone might need them, but the defaults were chosen because they sounded best on this setup. Leave them until you have a reason not to. The genuinely useful extras are speed (0.25–4x, default 1.0) and duration (0 = let the model estimate from the text).

How cloning actually works

OmniVoice is a masked diffusion language model, so generation is iterative: it starts from audio-shaped noise over a masked token sequence and refines it over num_step passes, with cfg pulling it harder toward the conditioning and t_shift nudging the mask schedule. There's no training involved - zero-shot means it learns the voice's patterns from the reference clip on the spot. Community testing puts it around 5GB of VRAM with cloning active.

The one thing people get wrong: the reference clip's quality is the ceiling. A flat three-second clip gives you a flat clone, and users who've pushed it report that 10 seconds is the usable minimum while 30–60 seconds with emotional variation dramatically improves expressiveness. If your clone sounds robotic, make the reference longer and livelier before you touch the sliders.

The real troubleshooting list

• "requires ref_text" - you passed audio without its transcript. It's the top user error.
• Non-finite or empty audio - the node itself suggests the fixes: it already runs fp32 (that's the default), so try disabling postprocess_output or using a cleaner/longer reference clip. Sometimes a stale run just needs a fresh generation.
• Install problems - this pack pins transformers==5.3.0, and the README's warning is in caps for a reason: it will not run on the 4.x many other packs force. pip list | grep transformer to check. This dependency war is the default failure mode of audio in ComfyUI, not a you problem.

One honest caveat: a few people find OmniVoice's emotion tags (the [laughter]-style markers from the upstream model) hit-or-miss, and opinions vary on how closely clones match the source. If the target is a dead-on impersonation of a very specific voice, treat this as "very good, sometimes uncanny," not magic.