Fish S2 Multi-Speaker TTS

The single-speaker Fish S2 nodes are already good. Fish S2 Multi-Speaker TTS is the one that makes you feel like you're cheating: it synthesizes a full conversation between up to ten cloned voices in one pass - two people bantering, a news anchors panel, a fake podcast interview - without you splicing audio or queueing ten separate generations.

You connect a reference clip for each participant, write the dialogue with [speaker_1]:-style markers, and the node returns one continuous audio file where the voices trade lines naturally. It builds on Fish S2 Pro's zero-shot cloning, so the "voices" are whoever you feed in - a couple of 10-second samples each and you've got your cast.

How it actually works

Worth knowing, because it sets expectations: the node doesn't generate the whole conversation in one autoregressive sweep. It parses your script into individual turns, generates each line separately using that speaker's reference audio (seeded seed + line_index so each line varies a bit), then concatenates the results with pause_after_speaker seconds of silence between turns - 0.4s by default, which reads as natural pacing for most dialogue. That's also why the node requires every speaker to have a reference clip connected: it'll throw a clear Reference audio required for all speakers error if one's missing.

You don't need to hunt the model's native <|speaker:0|> syntax either - the node converts your friendly [speaker_1]: text into the model format for you.

Inputs that matter

• num_speakers - 2 to 10. This is the fun part: it's a dynamic input. Change it and the node shows or hides exactly that many speaker_N_audio and speaker_N_ref_text fields on the panel. No fishing for the right dropdown.
• speaker_N_audio - the reference clip for each person (5–30s each).
• speaker_N_ref_text - optional transcript per speaker; recommended for stability, same reasoning as the voice-clone node.
• text - the script, one line per turn:

  [speaker_1]: Hello, I'm speaker one.
  [speaker_2]: And I'm speaker two!
  

  Inline [laugh] / [whisper] tags still work per line.
• pause_after_speaker - seconds of silence between turns (0–2). 0.4 is a good default; nudge it up for a more ponderous back-and-forth, down for rapid-fire comedy.

Everything else (model_path, language, precision, attention, temperature) is shared with the base Fish S2 TTS node, and the same guidance applies - leave them alone until you know why you're changing them.

Output is a single audio (AUDIO) track, ready for SaveAudio or a lip-sync model.

Install and gotchas

Same pack, same story: ComfyUI Manager → search "FishAudioS2" → install → restart, or clone https://github.com/Saganaki22/ComfyUI-FishAudioS2.git into custom_nodes/. Deps auto-install on startup; models auto-download to ComfyUI/models/fishaudioS2/ on first use. Watch the shared pack traps: don't pip install fish-speech (bundled, and it'll downgrade PyTorch), restart once if nodes fail to load, and use the BNB NF4/INT8 model options to squeeze this onto a 16GB card.

Two things that bite specifically on this node: the dynamic num_speakers inputs need a recent-ish ComfyUI (there's a legacy fallback if yours is old, but the dynamic version is much nicer), and every speaker must have a reference connected or the run fails on the first missing one. Also remember the Fish Audio Research license - non-commercial. Free for your projects, not for a product.

The one honest caveat: per-line generation means there's no cross-speaker prosody modeling - characters don't interrupt each other or talk over one another. For scripted, turn-based dialogue it's great and way faster than building it by hand. If you need genuine back-and-forth interruptions, that's a different (and heavier) problem entirely.