🔬 Spectral Processor

Normal audio processing happens sample by sample. Spectral processing is different: the audio gets transformed into the frequency domain, operated on there, and transformed back. That's how you get effects that sound like the audio's guts are being rearranged - and it's exactly what SpectralProcessor does, with ten modes. It's the most sci-fi node in the pack.

What it does

You feed in audio and get back processed_audio (plus, implicitly, whatever the FFT analysis says - the node is built on the same frame-based FFT engine as the analyzer nodes). The spectral_mode dropdown picks the operation:

• enhance / suppress - boost or cut a chosen frequency region (set by frequency_range_low/high).
• shift - move a band of frequencies up or down, the classic "spectral displacement" effect that makes things sound like a radio tuning through stations.
• morph - blend the spectrum toward something else, good for transitions and alien textures.
• gate - silence bins below a threshold, i.e., spectral noise reduction.
• compress - tame the loudest bins, spectral compression.
• chaos - scramble the spectral bins; pure digital weirdness.
• phase - mess with phase relationships between bins, which audibly smears and widens things.
• vocoder - the classic talkbox effect; use the frequency ranges as the filter banks.
• freeze - hold the current spectrum in place, producing an endless shimmering drone from any moment of the audio.

How it works

The pipeline is the standard STFT dance: audio is sliced into overlapping frames of fft_size (512–8192), windowed with window_type, FFT'd, the chosen mode operates on the magnitude/phase of each bin, then the inverse FFT and overlap-add reconstruct the time-domain signal. overlap_factor (0.25–0.95) controls the frame overlap - more overlap means smoother, more artifact-free results at a higher compute cost. The optional phase_randomization (0–1) scrambles phase on the way out, which makes things sound grainier and more "spectral" in the hazy way, and wet_dry_mix blends the processed result with the original.

intensity (0–2) is the global strength of the effect, and the frequency_range_low/high controls scope the effect to a band - crucial, because running full-spectrum chaos on everything gets old fast.

The inputs that matter

• spectral_mode - the identity. freeze and shift are the two you'll demo for friends.
• frequency_range_low / frequency_range_high - target a band and the effect becomes musical instead of a mess.
• intensity - how much of the effect you hear.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/eg0pr0xy/comfyui_noisegen.git
cd comfyui_noisegen
pip install -r requirements.txt

(Windows embedded Python: ComfyUI\python_embeded\python.exe -m pip install -r requirements.txt.) Or ComfyUI Manager → search "NoiseGen" / eg0pr0xy. No model files, no keys.

Gotchas

The pack rule applies - pip install is mandatory, scipy is imported at load and takes the whole pack with it if missing. Bigger fft_size values give finer frequency resolution but worse time resolution: on percussive material a large FFT smears transients into mush, so match the FFT size to the material (2048 is a good default for most noise). And phase_randomization at high values turns almost anything into pure texture - great for noise, wrong for anything you want to stay recognizable.