Random/Fixed Audio Picker (Creepybits)

The name is a promise: feed this node one long audio clip and it hands back a bite of it. Leave the mode on "random" and you get a different chunk every run - great for variation when you're building audio-to-prompt pipelines. Give it a fixed start time and it becomes a precision trimmer that can carve out "the ten seconds starting at the 1:30 mark" without you opening an editor. Two tools, one node, zero fuss.

It comes from the Creepybits suite (ComfyUI-Creepy_nodes), a grab-bag pack from the YouTuber behind the "Creepybits method" of ComfyUI installs. The audio node sits in their "Creepybits/Audio" category alongside tools that pull keywords out of audio and hand them to Gemini for prompt building - so this is the sampler that feeds those.

How it works

Under the hood it's just slicing a waveform tensor. ComfyUI passes audio around as a dict with a waveform and a sample_rate, and this node does the math on the sample count, cuts the slice, and hands back an AUDIO object with the same sample rate. Nothing fancy, which is exactly what you want from a utility.

Two modes, chosen by one value:

• start_time at its default of -1.0 means random mode. The node picks a start position anywhere the segment fits in the clip, so a 10-second segment out of a two-hour file gets pulled from a random spot every run.
• Any start_time of 0.0 or above means fixed mode. It starts exactly there - clamped so it can't run past the end of the clip.

The inputs that matter

• audio - the source clip. Wire in anything that outputs the AUDIO type (a ComfyUI audio loader, or the Load Audio node from VideoHelperSuite and friends).
• segment_length - seconds you want out, default 10, from 0.1 up to 600. If you ask for more than the clip contains, it just gives you the whole clip.
• start_time - the mode switch described above. This is the only one you'll actually touch.

The single output is audio, ready to feed a VAE/spectrogram node or one of the pack's audio-analysis nodes.

Installing it

This ships as part of ComfyUI-Creepy_nodes, so you get the whole suite at once:

cd ComfyUI/custom_nodes
git clone https://github.com/Creepybits/ComfyUI-Creepy_nodes
# then restart ComfyUI

Or just search "Creepy_nodes" in ComfyUI Manager and let it handle the rest. The pack pulls in numpy, opencv, requests, and google-genai - nothing model-shaped, no checkpoints to download.

Where it bites

Two gotchas worth knowing. First, random mode is genuinely random: there's no seed input, so you can't reproduce a specific pick. That's fine for inspiration, annoying if you thought you'd get consistency. Second, the module imports torchaudio at the top but the pack's requirements.txt never lists it - most ComfyUI installs that already handle audio have it, but if the whole pack fails to load after install, that import is the first thing to check.

If you just want a deterministic cut, set start_time and forget it. If you want surprise, leave it at -1 and let the node play roulette for you.