Audio AI Denoise (VoiceFixer)

VoiceFixer is the "restore the unfixable" node in this pack, and the author is refreshingly honest about when to use it. It's for audio that's genuinely wrecked - old movie dialogue, clipped recordings, tape hiss, the stuff that makes you wince. It is not for cleaning up freshly-generated speech, because it works by rebuilding the audio rather than filtering it, and that reconstruction can add its own artifacts to audio that was already fine. If your source is clean-ish AI speech, skip this node and reach for Audio AI Enhance (Resemble) instead - the README says so in so many words, and it's right.

How it works

VoiceFixer is a neural restoration pipeline (analysis module → synthesis module) that re-synthesizes speech, which is why it can do things simple EQ can't: it strips reverb, un-clips, and upscales bandwidth. The node runs the voicefixer library, which it auto-installs on first use. It also auto-manages the two model checkpoints - vf.ckpt (466MB) and model.ckpt-1490000_trimed.pt (135MB) - redirecting them into ComfyUI/models/voicefixer/ and symlinking the library's cache to that folder, so your models live where ComfyUI expects them instead of scattered in ~/.cache.

The inputs

• audio - the damaged track.
• mode - three gears:
  • Mode 0: Denoise + DeReverb + DeClip - full restoration, for the worst sources.
  • Mode 1: Denoise + DeReverb - the default middle ground.
  • Mode 2: Denoise Only - when you just want the hiss gone and don't trust re-synthesis.
• use_cuda - on by default; keep it on unless you enjoy waiting.
• splice_info (optional) - a passthrough input that comes back out untouched, so you can run this node inside a voice-conversion pipeline without losing the seam data. The output side mirrors it.

Outputs

audio (the restored AUDIO) and splice_info (passthrough, for the pipeline case).

Install and gotchas

The node itself is zero-config - install the pack, run it, and the library installs itself:

cd ComfyUI/custom_nodes
git clone https://github.com/havvk/ComfyUI_AIIA.git

Then the known pain points, straight from the README:

• Model downloads stall or die with PytorchStreamReader errors. The fix is manual placement - drop the two checkpoints into ComfyUI/models/voicefixer/ with the exact structure (analysis_module/checkpoints/vf.ckpt and synthesis_module/44100/model.ckpt-1490000_trimed.pt). Mirror links for both are in the README if HuggingFace is slow.
• First run is slow. It downloads the weights, then the model has to load - the pop-in install plus download can take a couple of minutes before anything happens.
• Artifacts on clean audio. This is the "don't" case, and it's the most common mistake. If your output develops a strange spectral shimmer, you fed it something VoiceFixer didn't need to fix.

When you'd actually reach for it

The honest workflow: try Audio Enhance first. If the source is genuinely bad - clipped, reverby, or old - switch to this, run Mode 1, and only escalate to Mode 0 if the reverb or clipping survives. It's a rescue tool, not a polish tool, and treating it as the latter is where people get burned.