SupaVoice Denoiser

Noisy audio is the silent killer of voice work. Whisper transcribes it wrong, and a voice-clone model trained on a hissy track learns the hiss. SupaVoice Denoiser sits between your audio source and your transcriber - or between your clip and your voice-clone training set - and strips the noise with one of two neural backends. It's the pack's only node that lives under audio/processing rather than SupaVoice/audio, and it earns the distinction.

How it works

Feed it any ComfyUI AUDIO and pick a backend:

• DeepFilterNet - fast, roughly 50MB of model, and excellent for speech. It processes internally at 48kHz and resamples back to your original rate. df_atten_db (default 100) controls the max attenuation in dB: 100 is full suppression, lower is gentler - handy when aggressive noise removal starts eating sibilance.
• Resemble Enhance - heavier, around 1GB, but it also upscales speech quality, which matters when the cleaned track is about to become a voice-clone reference. re_nfe (default 32) is the number of solver steps - higher is cleaner but slower. re_lambd (default 0.9) balances denoise vs enhance: 1.0 is full denoise, 0.0 is enhance-only. re_mode lets you pick denoise_only (faster) or denoise_and_enhance.

Inputs: audio and backend are required; everything else is optional and only applies to the matching backend. Output: a single audio (AUDIO) - mono, since both backends work in mono, at the resampled rate.

The pragmatic read: DeepFilterNet for speed, Resemble Enhance when the audio quality is going to directly affect clone fidelity and you have the patience and the 1GB download.

Installing it - read the requirements before you blame the node

Pack install is the usual:

cd ComfyUI/custom_nodes
git clone https://github.com/mexxmillion/ComfyUI-SupaVoice-Transcriber

Restart ComfyUI. Here's the gotcha that trips everyone: neither denoiser backend is in the pack's default requirements. The requirements.txt lists both as commented-out options. If you run the node and it throws ImportError, that's expected - install a backend yourself:

pip install deepfilternet       # ~50MB, the fast one
# or
pip install resemble-enhance    # ~1GB, the quality one

The node lazily loads whichever backend you pick on first run, so you can install one, use it, and add the other later without touching anything else. Don't install both into a crowded ComfyUI env unless you enjoy dependency roulette - the ecosystem's shared-environment problem is real (see the ComfyUI custom node ecosystem docs), and these are two more audio stacks fighting for the same site-packages. Also note the node calls torchaudio (or falls back to scipy) to resample, so if either of those is missing in your env, a resampling run is where it'll surface.

Use it after you've loaded/trimmed audio and before the transcription node - the order is exactly what it looks like, and it's the difference between a label set you trust and one you quietly don't.