⚙️ IndexTTS-2 Engine

IndexTTS-2 is the engine you pick when you care about how a line is delivered, not just what it says. Most TTS engines give you a voice and maybe a vague "expressiveness" slider. IndexTTS-2 lets you steer emotion directly - eight named emotions you can dial by hand, or an audio reference, or AI-analyzed emotion from the text itself, blended together in the model's conditioning space. For dialogue, character work, and anything where a flat read won't cut it, it's the standout in this pack. This node configures it and outputs a TTS_ENGINE for 🎤 TTS Text or 📺 TTS SRT.

It's a mid-size model (~4.7GB), covers English, Chinese, and Japanese, and clones a voice from a reference clip.

How it works

IndexTTS-2 conditions generation on both a speaker reference (the voice) and an emotion reference (the feeling), kept on separate paths so you can mix and match. That separation is the trick: you can clone Alice's voice but drive it happy, angry, or melancholic without needing a happy/angry/melancholic recording of Alice. The emotion side comes in through two optional inputs on this node, and emotion_alpha scales how much of it actually gets applied.

The inputs and outputs that matter

For a beginner, most of the required widgets are sampler internals you can ignore. The ones that matter:

• emotion_alpha (0–2, default 1) - the master emotion intensity. 0 is neutral, 1 is normal, 2 is maximum drama. This is the single most useful dial on the node. If emotion sounds too strong or too weak, this is what you turn.
• temperature (default 0.8), top_p (default 0.8), top_k (default 30) - standard sampling controls; defaults are fine.
• repetition_penalty (default 10) - note this default is unusually high on purpose, to stop the model looping. Don't casually drop it to 1.

Then two optional inputs that are the whole point of IndexTTS-2:

• emotion_control - takes an EMOTION_CONTROL from the 🌈 IndexTTS-2 Emotion Vectors node (manual 8-emotion sliders) or a text-emotion source. This is the "set the feeling by hand" path.
• emotion_audio - takes any audio clip as an emotion reference; the model copies its emotional tone. You can feed a Character Voices opt_narrator here for per-character emotion references. Both can be used at once and get blended.

Advanced toggles like use_fp16 (on by default, saves VRAM), low_vram, use_cuda_kernel, and use_deepspeed are performance switches - leave them until you hit a memory wall, then flip low_vram. Output is a single TTS_engine.

Installing it

Comes with TTS Audio Suite. Easiest is ComfyUI Manager: search TTS Audio Suite, install, restart - Manager runs the pack's install.py, which 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 just a requirements pip-install). Linux: sudo apt-get install portaudio19-dev libsamplerate0-dev first. The IndexTTS-2 model (~4.7GB, emotion components included) auto-downloads on first use into ComfyUI/models/TTS/IndexTTS/. There's a "🌈 IndexTTS-2 Integration" example workflow in the repo that wires the emotion inputs for you.

Common issues

• Emotion isn't showing up. Check emotion_alpha first - at 0 you've muted it entirely. Then check that something is actually connected to emotion_control or emotion_audio; the node happily generates neutral speech with nothing plugged in.
• Over-dramatic, warbly output. emotion_alpha too high, or an emotion audio reference that's very intense. Pull alpha back toward 1.
• Looping or repeated words. That's what the high default repetition_penalty guards against - if you lowered it, put it back up.
• VRAM errors. It's a 4.7GB model plus emotion components. Enable low_vram and keep use_fp16 on. Per-character emotion via [Character:emotion_ref] tags is powerful but multiplies the references the model juggles.