RunningHub VoxCPM Multi-Speaker (Dynamic Audio)

This is the smarter sibling of the fixed five-slot Multi-Speaker node, and the answer to the complaint that node leaves you staring at empty sockets. RunningHub VoxCPM Multi-Speaker (Dynamic Audio) uses the same tagged [spk1]...[spk2] script format but lets reference-audio inputs appear and disappear as you actually connect them - it starts with two, and adds the next one every time all visible inputs are connected. No fixed upper limit, so a 12-voice radio drama is just a matter of wiring more clips.

The mechanism is worth knowing because it's the one thing in this pack that's a little magical. A frontend extension script (the pack ships it in js/voxcpm_audio_list.js) watches the node's connections and renames, adds, and trims audio_1, audio_2, ... inputs on the fly. That's browser-side behavior, which means two practical gotchas: it needs a page refresh after the pack updates to pick up the script, and the audio_N slot order maps to spkN at execution time - so audio_10 really does feed [spk10] in your script.

What the node actually asks for

• script - the tagged dialogue, same format as the fixed version.
• speaker_controls - this is the big difference. Instead of one control_N widget per speaker, you write all descriptions in a single multiline field using the tags themselves:

  [spk1]Sichuan accent
  [spk2]Adult female, northeastern accent
  [spk5]Narrator, mature male
  

• reference_texts - optional tagged exact transcripts for each clip, e.g. [spk1]Reference words. Blank falls back to VoxCPM2's reference-only conditioning, which is safe and usually fine.
• audio_1, audio_2, ... - the dynamic reference clips. Leave one empty and it stays a designed voice; connect them all and a new slot appears.

The rest is the standard set: cfg_value (2.0), inference_steps (10), seed, normalize_text, denoise_reference, max_len (4096), retry_badcase (on). Output is one audio wire with the whole dialogue.

Setup

Identical to the rest of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/RH-RunningHub/ComfyUI_RH_VoxCPM.git
cd ComfyUI_RH_VoxCPM && pip install -r requirements.txt
hf download openbmb/VoxCPM2 --local-dir ComfyUI/models/voxcpm/VoxCPM2

Restart ComfyUI, then - and this is the part people skip - hard-refresh the browser tab. The dynamic inputs are driven by a frontend extension, and a stale page is the #1 reason this node looks broken right after install.

Traps to sidestep

The auto-growth is the whole feature and the whole failure mode. If you connect audio_1 and audio_2 and no audio_3 appears, the JS didn't load - refresh. If you then connect audio clips in a different order than the tags in your script, the voice-to-speaker mapping goes sideways because mapping is positional, not by tag content. And if you disconnect an audio input, the script trims trailing inputs, which can renumber slots mid-workflow - don't build a huge multi-speaker graph on top of a live-growing socket list. For a fixed cast of five or fewer, the plain Multi-Speaker node is honestly the more predictable choice; this one earns its keep when your speaker count varies between runs or climbs past five.