OmniVoice Voice Preset

OmniVoice Voice Preset is the "give me a voice right now" node. OmniVoice is a cloning model, which means it needs a reference audio clip to clone - and if you don't have one, you're stuck. This node solves that by shipping nine pre-fetched reference voices, all bundled with their transcripts, ready to wire straight into OmniVoice Generate. Drop the node, pick a preset, connect its two outputs, and you're generating speech with a working voice in under a minute.

It's the fastest on-ramp to the whole pack, and it's what the included example workflow (workflows/voice_cloning.json) uses. The presets are a fun mix, leaning on the same celebrity-adjacent and character voices that drove the original source clips: Shadowheart, Rick Sanchez, Stewie Griffin, Harvey Keitel, Conan O'Brien, plus archetype voices like "Podcast host" and "Nature." Seven come from Chatterbox demo clips, one from F5-TTS - which is a nice reminder that these are just reference audio, not model-trained voices; OmniVoice clones them at runtime like any other clip.

How it works

Each preset is a URL and a transcript pair baked into the pack. On first use, the node downloads the clip once and caches it to ComfyUI/models/omnivoice/presets/, then loads it as a ComfyUI AUDIO value. The transcript (the exact ref_text that makes cloning work well) comes out the second output. Because the transcripts are pre-filled, these presets sidestep the #1 cloning pitfall - missing transcription - without you doing anything.

Outputs: ref_audio (AUDIO) and ref_text (STRING), which wire straight into the matching inputs on OmniVoice Generate. That's the whole interface; there's one dropdown (preset) and nothing else to configure.

Adding your own presets

This is the sleeper feature: you're not stuck with the nine. Drop any audio file (wav, flac, mp3, ogg, m4a) into ComfyUI/models/omnivoice/presets/, add a same-name .txt file alongside with its transcript, restart ComfyUI, and it appears in the dropdown labeled (local). It's a genuinely handy way to keep your go-to reference voices in the same UI as the built-ins - and since the node only scans for files that aren't cached built-ins, your additions won't collide with the defaults.

Installing it

Standard for the pack - ComfyUI Manager (search "OmniVoice") or:

cd ComfyUI/custom_nodes
git clone https://github.com/ethanfel/ComfyUI-Omnivoice.git

then restart. The presets download lazily on first use, so don't be surprised by a brief hang the first time you pick one - subsequent runs are instant.

Common issues

• First generation is slow - did it break? No. It's the model downloading (several GB to models/omnivoice/) plus the preset fetching. Give it time once and it's cached.
• Presets aren't in the dropdown after adding files. The scan happens when the node builds its input list, which is at load time. Restart ComfyUI (and refresh the browser tab) after dropping files in.
• The voice sounds like the source, not what you hoped. These are real character voices - Rick Sanchez sounds like Rick Sanchez. That's the point. If you want a generic voice, the Podcast host and Old Hollywood presets are safer bets than the IP-adjacent ones.
• Transcript mismatch if you hand-edit. If you write your own .txt transcript, keep it accurate to the clip - the whole preset system leans on ref_text being right, and a wrong transcript is worse than an empty one.

One honest note: these are reference clips of real people's voices, and the IP-flavored ones (Rick, Stewie, Shadowheart) are for tinkering, not for shipping a commercial product. For your own project, the clean move is a preset for prototyping, then a real (consented) reference voice you can own. The model doesn't care where the clip came from; your liability profile does.