OmniVoice Longform TTS

If you install exactly one node from the OmniVoice pack, make it this one. OmniVoiceLongformTTS is the do-everything text-to-speech node: throw a whole paragraph - or a whole script - at it and it chunk the text at sentence boundaries, synthesize each piece, and hand you back one clean AUDIO track. Plug in a reference clip and it turns into a voice cloner. Leave the reference disconnected and it just picks a voice for you. No API, no key, fully local, and it claims 600+ languages - which is a much bigger deal than it sounds when most open TTS is English-first.

The engine underneath is OmniVoice by k2-fsa: a diffusion language model built on a Qwen3 backbone (the README is refreshingly honest about that in its attention-backend section) that generates speech token-by-token rather than in one shot. That's what makes it both fast - RTF down around 0.025, meaning 40x real-time - and multilingual. It landed in April 2026 to a genuinely split reception: half the thread loved it ("the model is great and fast"), the other half compared it to VibeVoice or Chatterbox and shrugged. My take: it's the best "one pack, many languages, no training" option in ComfyUI right now, and the longform node is the cleanest way to exercise it.

The input list is long, but a beginner only needs a handful. model is a dropdown - OmniVoice (full fp32, ~4GB) or OmniVoice-bf16 (~2GB), both auto-download to ComfyUI/models/omnivoice/ on first run. text is your script; it supports inline non-verbal tags like [laughter], [sigh], [question-en], which are nice when they work - fair warning from the community: they're hit or miss, and sometimes the model reads them aloud or ignores them entirely. words_per_chunk (default 100) is the chunking knob; 0 disables it. ref_audio and ref_text are optional - connect a 3–15 second clip and it becomes a voice-clone job; leave ref_text empty and Whisper transcribes the clip for you. instruct lets you force dialect and style, but only from a fixed list (male, indian accent, whisper, and a bunch of Chinese dialects like 河南话), so don't free-type there.

The sampler-ish knobs are shared across every OmniVoice node: steps (16 = fast, 32 = balanced, 64 = best), guidance_scale (2.0 default), speed, and the three temperature/penalty knobs you can safely ignore on day one. keep_model_loaded stays on - it offloads the model to CPU between runs rather than fully unloading. The output is a single audio (AUDIO) that wires straight into a save-audio node, or into the next stage of a talking-head pipeline if you're building a lip-sync workflow.

Install is ComfyUI Manager - search "OmniVoice" - or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Saganaki22/ComfyUI-OmniVoice-TTS.git
cd ComfyUI-OmniVoice-TTS
python install.py

One real gotcha, and the pack's README is upfront about it: the omnivoice pip package pins torch==2.8.*, which can downgrade your PyTorch to CPU-only and kill ComfyUI's GPU. The installer gets around it with --no-deps, but if your PyTorch ever goes weird, check the pack's pytorch_compatibility_matrix.md. The other classic failure: only the Whisper Loader node shows up and the TTS nodes are missing - that's omnivoice failing to import, usually because transformers is too old (pip install "transformers>=5.3.0" plus soxr), followed by a full ComfyUI restart. Expect roughly 5GB VRAM with the bf16 model and cloning enabled, per community measurements.