Voice Emotion Audio

IndexTTS-2 lets you control the emotion behind a cloned voice three ways: a text description, an emotion vector, or - this node - a second audio clip that demonstrates the emotion. Voice Emotion Audio is the one to pick when you can't describe the feeling in words but you can point at a recording of someone doing it. Think of it as stealing the delivery from one clip and grafting it onto a different speaker's voice.

It's the most literal of the emotion inputs, and for a lot of use cases it's the most controllable. Want the cloned voice to sound like it's about to cry? Find a clip of someone on the verge of tears and use that as the emotional reference. The model transfers the prosody, not the voice - the timbre still comes from your speaker reference.

What you set

• reference_audio - the voice to clone. This is the identity; it never changes.
• emotion_ref_audio - the emotional delivery to borrow. This is the mood; it's layered on top of the speaker.
• emo_weight - how strongly the emotion lands, from 0.0 to 1.6, defaulting at 0.8. Crank it up and the delivery gets theatrical; dial it low and it's barely there. This is the control you'll actually tune between takes.
• voice_name - optional label used in multi-voice scripts. Leave it blank and the voice is auto-named s1, s2, ... in merge order.

The output is a single emotions object (EASY_INDEXTTS_EMOTIONS) that flows into IndexTTS Generate, IndexTTS Generate Simple, or Merge Voice Emotions for multi-voice scenes. Every emotion node produces this same wire type, which is the point - the Generate node treats them all uniformly, no matter which one you used to make them.

Honest advice

This is the most fiddly of the three emotion inputs, because it inherits the classic TTS trap from the reference-cloning world: quiet or flat reference clips produce weak results. A loud, dynamic emotion clip works dramatically better than a whisper-quiet one - same rule that applies to the voice reference itself. The emo_weight slider is your friend here; when the borrowed emotion feels faint, push it toward 1.2–1.4 before you go re-recording anything.

Also worth knowing: emotion audio outranks the other emotion modes internally, so if a voice has both an emotion-audio clip and a vector or text emotion, the audio wins. Set up a clean pair of reference clips once - a dry voice sample and a juicy emotion sample - and this node becomes a one-drag habit.