FL VoxCPM Audio Crop

The dumbest node in this pack might save your clone. Voice cloning wants a clean, short reference - a few seconds of a voice with no music, no coughs, no dead air. If you drop a whole podcast episode into reference_audio, the model doesn't know which voice to learn and the retry guard has a field day. This node is the trimmer: you feed it audio, give it a start and end time, and it hands back just that slice.

It's part of filliptm/ComfyUI-FL-VoxCPM, the same pack as the VoxCPM TTS nodes, and it speaks the ComfyUI AUDIO format natively - so it chains straight into prompt_audio on the V1 TTS node or reference_audio on V2.

How it works

Plain waveform surgery. You give it three things:

• audio - any AUDIO socket (Load Audio, a previous node, whatever).
• start_time - 0:00 by default. Written as MM:SS, or just seconds like 12.
• end_time - 1:00 by default. Same format.

It parses the times, converts them to sample frames at the audio's sample rate, and slices the waveform tensor. Output is a single audio (AUDIO) socket. There's nothing clever going on, which is the point - it's a utility you string into the middle of a workflow when your source clip isn't already production-clean.

Inputs that matter

The two time strings are the whole node. A couple of practical notes: the times are validated, so if start_time >= end_time you'll get an error rather than a silent empty clip, and anything out of range gets clamped to the audio length rather than crashing.

Installation

One install covers the whole pack:

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

Or Manager → search "FL VoxCPM". No model download for this one - it's pure tensor math, so it's the cheapest node in the pack.

When you'd actually use it

Two workflows, really. Voice cloning: trim your reference to the cleanest 5–15 seconds before it hits the TTS node, and transcribe that exact slice (FL VoxCPM Transcribe makes that easy). Dataset building: when you're feeding clips to the Dataset Maker for LoRA training, each sample works best if it's a tight, complete sentence - crop the dead time out of each file first and your training set gets dramatically better for free.

If your audio is already cut exactly the way you want it, you don't need this node. But the moment a clone sounds "off" and you're not sure why, the culprit is usually a sloppy reference - and this is the two-second fix.