VoxCPM Audio Converter

ComfyUI's own LoadAudio and the audio types that TTS nodes expect are not the same shape, and this is the node that admits it. VoxCPM Audio Converter takes any ComfyUI AUDIO and rebuilds it into exactly what the VoxCPM nodes want: mono, float32, laid out as [1, 1, samples]. One input, one output, same type on both sides - it's a pass-through that normalizes.

The pack's README is unusually blunt about when to use it: "Place between LoadAudio and VoxCPM nodes if you get 'audio too short' errors." That error is almost never a real duration problem. It's a channel or dtype problem - stereo audio that the node can't flatten the way it expects, an int16 waveform, a shape that confuses the length check. The converter squashes all of that: if the waveform is multi-dimensional it averages channels, forces float32, squeezes to 1D, and rebuilds the AUDIO dict with an explicit integer sample rate.

The one input, the one output

• audio (AUDIO) - whatever ComfyUI's LoadAudio (or another audio node) handed you.
• audio (AUDIO) - the same audio, renormalized for VoxCPM consumption.

There's genuinely nothing to configure. It even prints a little diagnostic to the console showing the input shape before and the output shape after, which is more helpful than most nodes this size manage.

When you actually need it

If you load reference audio through this pack's own VoxCPM Load Audio node, you probably don't need the converter - that loader already returns normalized mono float32, which is why the README calls it the recommended path. You want the converter when you're using ComfyUI's stock LoadAudio (or dragging in audio from another node pack), because that's where the shape mismatch shows up. Keep it in the graph as a matter of habit and the "audio too short" family of errors mostly stops existing.

Installing it

It ships in the ComfyUI-DN-VoxCPM pack, so:

cd ComfyUI/custom_nodes/
git clone https://github.com/DekaNear/ComfyUI-DN-VoxCPM.git
cd ComfyUI-DN-VoxCPM
pip install -r requirements.txt

Restart ComfyUI (or install via Manager by searching "DN VoxCPM"). The node itself is pure numpy, but the pack's requirements.txt drags in voxcpm and funasr, so you pay the heavy install for the rest of the family even if you only came for this one.

Where people get burned

There's no real trap here, just a diagnostic. If the converter itself throws "audio is too short" on genuinely short audio (under 100 samples), your source is broken - a near-empty recording, a failed download, a zero-length export. Fix the source, not the node. Otherwise, treat it as the cheapest troubleshooting step in the pack: audio weirdness downstream of LoadAudio? Insert converter, rerun.