OmniVoice Multi-Speaker TTS

The demo that sells this pack. OmniVoiceMultiSpeakerTTS generates a whole conversation in one pass - two, three, up to ten different cloned voices trading lines, with proper pauses between speakers - and outputs a single AUDIO file. No stitching clips together in an editor, no per-speaker generation and concatenation, no alignment pain. If you've ever tried to fake a dialogue with a single TTS voice, this node is the reason to stop.

The trick is dead simple on your side: mark each line with a [Speaker_N]: tag and wire each speaker's reference audio to its numbered input.

[Speaker_1]: Hello, I'm speaker one.
[Speaker_2]: And I'm speaker two!
[Speaker_1]: Nice to meet you!

num_speakers (2–10) controls how many speaker inputs appear. On ComfyUI 0.8.1+ these are dynamic inputs that show and hide as you change the count; on older ComfyUI the node falls back to a legacy mode where the inputs are always present. For each speaker you get three optional inputs: speaker_N_audio (their 3–15 second reference clip - required if that speaker is going to sound like anyone in particular), speaker_N_ref_text (transcript; empty = auto-transcribed with Whisper), and speaker_N_instruct (a per-speaker dialect/style push, from the model's fixed list like male, elderly, 河南话). If you leave a speaker's audio unconnected, OmniVoice will still render their lines with a synthetic voice - handy for that one extra character you don't have a reference for.

Under the hood it's the same zero-shot diffusion TTS as the rest of the pack, just parsing your tagged script, generating each speaker's lines against their own voice tokens, and layering in pause_between_speakers (0.3s default) so the back-and-forth reads as a real conversation rather than clipped together. Generation runs per speaker, so steps (32 default) and speed apply to all of them; the other knobs - guidance_scale, t_shift, the temperature trio - match the other nodes. The output is one audio (AUDIO) track with everyone already in place.

Quality follows the same rules as the single-speaker clone node, and they're worth repeating because a bad reference ruins a whole dialogue: use clear, 3–15 second samples, and if you want an argument that actually escalates, pick references with emotional range - the model's expression stays close to what its source clips contain. The non-verbal tags ([laughter], [sigh]) work in multi-speaker text too, with the usual caveat that they're sometimes ignored or spoken aloud.

Install is pack-standard: search "OmniVoice" in ComfyUI Manager, or git clone https://github.com/Saganaki22/ComfyUI-OmniVoice-TTS and run python install.py. The shared gotchas all apply: the installer uses --no-deps on omnivoice specifically because its torch==2.8.* pin can downgrade your PyTorch and break GPU acceleration - don't install it any other way. If the node is red or missing while the Whisper Loader works, that's the omnivoice import failing, almost always a transformers too old for the model's tokenizer (5.3+ needed) or a missing soxr, fixed with a pip install and a full restart. VRAM with bf16 lands around 4–6GB depending on how many speakers are live at once - keep_model_loaded offloads to CPU between runs, so it settles back down.