Sonder Reference Audio Bridge

Most reference nodes feed images, because most of the time that's what you want to stay consistent: a face, a location, an outfit. But the video models of 2026 increasingly condition on audio too - a voice identity to carry across a scene, an ambience bed, a vocal performance to keep. The Sonder Reference Audio Bridge is the sibling that handles that side of the ComfyUI-Sonder-Editor reference system, taking trimmed audio members off the timeline and handing them to your model as real AUDIO tensors.

Same shape as its image counterpart: a Sonder Reference Selector upstream resolves which lanes are effective for the current render window, and this bridge decodes what it resolved. One lane per media kind means audio lives on its own lanes - you can't drop a voice clip on an image lane - and the bridge only sees homogeneous audio reference sets. It concatenates lanes in lane-index order and emits up to sixteen trimmed audio members, one per output socket, in staged member order. The trim is set in the editor (waveform trim with preview), so the model gets exactly the snippet you cut, not the whole source file.

The inputs and outputs

Wire in the reference_set from a Sonder Reference Selector - that's the whole required input. Outputs are a01 through a16, all AUDIO, ordered by staged member. If a lane's recipe is an audio recipe, each staged member maps to its own socket; the built-in recipes include LTX ID-LoRA Voice Identity (up to 16 members) and MiniMax H3 Standalone Audio (3), so which model you're driving decides how many sockets actually fill.

The one optional input, unused_slots, is where beginners trip. Default is placeholder, which emits a one-second silent stereo track for every socket the recipe doesn't drive. That sounds wasteful, and it is - unless the node wired downstream has a required audio input, in which case it needs something on that socket. Switch it to nothing when the consumer's inputs are optional: a silent placeholder fed to an optional input can be read as real content, so an empty value is the honest signal that no reference is staged for that slot. The canvas marks a placeholder socket wired into a proven-required input as (unused · required input), so the mismatch shows before you burn a render.

Why you'd actually reach for it

If your model does audio at all - LTX 2.x with audio, MiniMax H3 with native stereo, and the like - this is how a character's voice follows them scene to scene without you loading a WAV by hand per shot. Stage the voice member on a lane, scope it over the frames where the character speaks, and the editor feeds the right trim to the right render window automatically. If you're only generating silent clips, you can skip this node entirely; audio references only matter for models that consume them.

Install and gotchas

It ships in the Sonder Editor pack alongside the Selector and the image/prompt bridges - install once, use all of them:

cd ComfyUI/custom_nodes
git clone https://github.com/SonderSaid/ComfyUI-Sonder-Editor.git
cd ComfyUI-Sonder-Editor
pip install -r requirements.txt

Restart ComfyUI afterwards. ComfyUI-Manager can do it too once the pack reaches the Registry - it's not there yet, so the clone above is the reliable route right now. The usual pack-wide caveats apply and are worth repeating because they're the ones that actually waste your time:

• Update ComfyUI first if nodes are missing. The Reference nodes run on the newer V3 node API and won't register on older builds.
• Audio decode needs ffmpeg. imageio-ffmpeg bundles a usable fallback, but a system ffmpeg on your PATH gives the widest format coverage - the README recommends it.
• torch is not installed by the pack, on purpose. Audio features may want torchaudio; if it's missing, install the build matching your existing torch/CUDA rather than letting any auto-resolver drop a mismatched wheel.
• Watch the OpenCV variant. opencv-python-headless is the deliberate choice here; the full opencv-python from another pack will fight it and break cv2 imports.

If you hear nothing back from a voice reference, the first things to check are unused_slots (silent placeholders can hide in optional inputs) and whether the audio lane actually has an item effective for your render window - the bridge only decodes what the Selector resolved.