IndexTTS 2.5 角色音色库 · T8star-Aix

One voice profile is a cast member. This node is the casting call: it takes 1 to 16 角色音色 (voice profiles) and bundles them into a single T8_INDEXTTS25_ROLE_LIBRARY that the multi-role/SRT generation node reads. If you're doing a two-character dialogue, a full audiobook cast, or anything with more than one voice in one audio file, this is the collection point in the middle of the graph.

How it works

The voices input is an autogrowing socket (COMFY_AUTOGROW_V3) - drag more connections onto it and it grows to fit, up to 16. Each input takes a voice profile (role name + reference audio + language + optional emotion). The node validates on the way in and returns:

• 角色音色库 - the merged library, fed to the multi-role/SRT generation node.
• 角色列表 - a string you can read (or dump to a text preview) showing exactly which roles made the cast.

Two behaviors are worth knowing before they bite you. Duplicate role names are rejected at queue time - the node refuses to run if two profiles share a name, because downstream the script resolves 角色A by name and ambiguity is a bug you'd rather catch at the door. And despite the display name sometimes reading like a mixer, "merge" here means collect - it does not blend multiple characters' eight-dimensional emotion vectors into one new mood. If you want "sad 60% + angry 40%", that's a single emotion-control node setting two dimensions, not this node.

The alias that'll confuse your searches

The pack ships a second node, Merge Voice Emotions, that does exactly this same job under the community-familiar search name. Both output the identical T8_INDEXTTS25_ROLE_LIBRARY. Use whichever you remember; they're interchangeable. The README's multi-role diagram shows the intended shape:

情感控制A ──► 角色音色A ┐
参考音频A ──►            │
                         ├─► 角色音色库 ─► 多角色/SRT生成
情感控制B ──► 角色音色B │
参考音频B ──►            ┘

Each voice's saved emotion gets read per-line - character A's lines carry A's emotion, B's lines carry B's.

Practical notes

It's a pure data node: no model load, no synthesis, no VRAM. Ships with the pack, so install once (ComfyUI Manager search IndexTTS 2.5 · T8star-Aix, or clone the repo) and every node in the family arrives together. The genuinely useful output for debugging is 角色列表 - when the multi-role generator tells you a role has no connected voice, checking the cast list makes the missing name obvious in one glance.

One honest caveat: 16 voices is a lot of reference audio flowing through the graph, and this node doesn't do any quality policing. Garbage reference clips produce garbage cast members - the library faithfully packages whatever you feed it. Cast your voices well here and the downstream work is mostly writing the script.