⚙️ Fish Audio S2 Pro Engine

Fish Audio S2 Pro is the "I need real breadth and nuance" engine. 80+ languages, zero-shot voice cloning, native multi-speaker dialogue, and - the interesting part - free-form emotion and prosody tags you drop right into the text at the sub-word level. It's a large model (~8–10GB depending on checkpoint), so it's not the one you reach for to knock out a quick English clip, but when you want expressive, multilingual, conversational output it earns its footprint. This node configures it and outputs a TTS_ENGINE for 🎤 TTS Text or 📺 TTS SRT.

How it works

S2 Pro is an LLM-style TTS (hence the temperature/top_p/repetition_penalty/context_length controls). It clones a voice zero-shot from a reference and reads your text, honoring inline tags for emotion and prosody as it generates rather than as a separate editing pass. Because it's a big transformer, it exposes real production controls: precision, quantization to fit smaller GPUs, and torch compile for speed. The suite runs it in the main Transformers 5 environment and tears the process down to clear VRAM when you're done.

The inputs and outputs that matter

• model (s2-pro or s2-pro-fp8, default s2-pro) - the full BF16 checkpoint or a lighter community FP8 one. FP8 trades a little quality for a smaller memory footprint.
• temperature (0.1–1, default 0.8), top_p (default 0.8), repetition_penalty (default 1.1) - LLM sampling; defaults are tuned. Raise repetition_penalty if it loops.
• multi_speaker_mode (Native Multi-Speaker or Custom Character Switching, default Native) - Native does a single-pass dialogue with real conversational flow; Custom is the standard per-character generation with full tag support.
• language_prompting (Auto Inline Tag or Off, default Auto) - whether it auto-inserts language tags for you.

The performance/memory group is where you'll actually intervene on a modest GPU: quantization (none, bnb_int8, bnb_nf4) lets you quantize the official checkpoint on load to fit less VRAM; precision (bfloat16/float16) and compile (torch.compile, on by default for speed) round it out. native_chunk_length, max_new_tokens, and context_length govern how much text it handles at once. speaker2 is a wildcard input for a second dialogue voice. Output is one TTS_engine.

Installing it

Comes with TTS Audio Suite. Easiest: ComfyUI Manager → search TTS Audio Suite → install → restart, which runs the pack's install.py and handles the dependency conflicts (NumPy, librosa, s3tokenizer) and Python 3.13. Manual: cd ComfyUI/custom_nodes && git clone https://github.com/diodiogod/TTS-Audio-Suite.git, then python install.py from inside the folder with your ComfyUI venv active (run the script, not a bare requirements install). Linux: portaudio19-dev libsamplerate0-dev first. The model (~8–10GB) auto-downloads into ComfyUI/models/TTS/fish_audio_s2_pro/ on first use. There's a "Fish Audio S2 Pro Integration" example workflow demonstrating cloning, native dialogue, and inline control. Note the model ships under the Fish Audio Research License - check it before any commercial use.

Common issues

• Out of VRAM. It's a big model. Set quantization to bnb_nf4 (or bnb_int8), or use the s2-pro-fp8 checkpoint. Both cut memory at a small quality cost. The suite's process teardown on "Clear VRAM" also helps between runs.
• First run stalls, then a torch compile pause. The download is ~10GB, and with compile on the first generation eats extra time compiling the graph - subsequent runs are faster. Not a hang.
• Emotion tags not landing. S2 Pro reads free-form sub-word emotion/prosody tags in the text; if you're not seeing an effect, check you're using its tag syntax (see the pack's engine docs) and not another engine's.
• Overkill for simple English. If you just need a clean English clip, ChatterBox or F5 loads faster and lighter. Save S2 Pro for when you actually need the 80-language reach or the fine emotion control.
• License. Research License, not MIT like the pack code. Relevant if you're shipping something commercial.