Fish S2 Multi-Speaker Split TTS

Take the multi-speaker conversation node and add the one output it's missing for animation work, and you get this. Fish S2 Multi-Speaker Split TTS generates the same multi-voice dialogue as its sibling, but instead of handing you one mixed audio file it also gives you a separate track per speaker - each track holds that speaker's lines and digital silence everywhere else, all the same length as the combined audio. If you've ever tried to lip-sync a two-person scene and had to manually slice one wav into "who's talking now" chunks, you already know why that's worth a whole node.

This is the ComfyUI answer to the "generate voices, then drive mouths" pipeline the KB documents - the standard local talking-head chain is TTS → audio-driven avatar, and for multi-character scenes (Infinite Talk and friends) each mouth needs its own audio track to know when to move. The per-speaker isolation idea came from the community (@lazybuttalented suggested it), and it's the difference between "render a two-person conversation" and "render a two-person conversation where both faces move correctly."

How it works

Under the hood it's the multi-speaker node's generation loop - parse the script, generate each line with the right speaker's reference, pause between turns - plus a routing step. As each line finishes, it's appended to the speaking speaker's track while every other track gets a zero-padded segment of the same duration. The result: audio is the full conversation, and each speaker_N_audio output is an aligned, isolated stem. pause_after_speaker controls the silence between turns and applies to both the combined file and the tracks.

Inputs that matter

Inputs are identical to Fish S2 Multi-Speaker TTS, so everything you learned there carries over:

• num_speakers - 2 to 10, dynamic; raising it reveals more speaker_N_audio / speaker_N_ref_text inputs.
• speaker_N_audio - a 5–30s reference clip per voice. All required, or the run fails.
• text - dialogue in [speaker_1]:/[speaker_2]: format, inline emotion tags allowed.
• pause_after_speaker - silence between turns (0.4s default).

The outputs are the whole point: audio (the mix), plus speaker_1_audio through speaker_10_audio (AUDIO). Wire each speaker track to its own lip-sync node and keep audio as your master for the final render. You don't have to connect all ten - unused outputs are just empty tracks.

Shared pack knobs - model_path, precision, attention, temperature, keep_model_loaded - behave exactly as in the other Fish S2 nodes. Start on s2-pro-bnb-nf4 if you're under ~20GB of VRAM.

Install and gotchas

ComfyUI Manager → search "FishAudioS2" → install → restart, or git clone https://github.com/Saganaki22/ComfyUI-FishAudioS2.git into custom_nodes/. Dependencies auto-install at first startup; the model auto-downloads to ComfyUI/models/fishaudioS2/. Don't pip install fish-speech yourself (bundled, PyTorch-downgrade trap), and give it one restart before debugging load failures.

Realistic expectations, same as its sibling: this is turn-based dialogue with pauses, not overlapping conversation - each track is cleanly "this speaker, then silence," which is exactly what lip sync wants anyway. One tip for animation: if you're driving faces, keep pause_after_speaker at 0.3–0.4s so the "not speaking" moments read naturally on screen. And the usual license reminder - Fish Audio Research terms, non-commercial unless you buy a commercial license from Fish Audio.

If you're building multi-character talking-head videos in ComfyUI, this is the TTS node you want. It quietly removes the most fiddly manual step in the pipeline.