Split Audio

A one-node way to cut a clip at a single timestamp and get both halves back - useful for dataset work where you've found a natural break point (a pause, a sentence boundary from a transcript) and want the audio on either side of it, not just one cropped piece.

How it works

It splits the waveform at the point you specify, in seconds, and hands back both resulting pieces. There's only one audio output socket in the schema, not two - the pack's own README explains why that's not a contradiction: AUDIO in this toolkit "consists of waveform and sample rate," and the waveform tensor carries a batch dimension. Rather than two separate sockets, SDT_SplitAudio almost certainly returns both halves stacked along that batch dimension in a single AUDIO object - the same pattern the pack leans on elsewhere for handling multiple clips as one typed value.

The inputs and outputs that matter

• audio - the clip to split, required.
• second (default 0, step 0.001) - the timestamp to split at, millisecond precision.
• audio (output) - both resulting segments, batched into one AUDIO value per the pack's batch-dimension convention above.

How to install it

ComfyUI Manager: search ComfyUI-speech-dataset-toolkit, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/kale4eat/ComfyUI-speech-dataset-toolkit
cd ComfyUI-speech-dataset-toolkit
pip install torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt

No extra dependencies - standard torchaudio baseline.

Common issues & troubleshooting

Downstream node only seems to process one of the two halves. If a node further down the chain isn't batch-aware, it may only act on the first item in the batch dimension rather than both. Check whether you actually need both halves independently - if so, SDT_TrimAudio twice (once per side) with explicit start/end timestamps is the more predictable path than relying on batch handling downstream.

Not sure which half is which after splitting. Given the split point is the boundary, the first item in the resulting batch should be everything before second, the second everything after - but since this behavior isn't spelled out in the pack's docs, confirm it against your actual audio with SDT_PlayAudio before building a pipeline that depends on the ordering.

This is one of the more obscure nodes in an already niche pack - it has almost no visible usage or discussion anywhere online, so if you hit unexpected batch-dimension behavior, you're mostly on your own; reading the node's source directly in the cloned repo is the fastest way to confirm exactly what it returns.