kotoba-whisper List Segments

A small but necessary plumbing node. SDT_KotobaWhisperTranscribeShort and SDT_KotobaWhisperTranscribeLong both hand back their segments output as one bundled KOTOBA_WHISPER_SEGMENTS object - not something you can pull individual start/end/text values out of directly. This node cracks that bundle open into a list of individual KOTOBA_WHISPER_SEGMENT items you can actually iterate over.

How it works

It doesn't do any transcription or processing itself - it takes the bundled segments object and re-exposes it as a list output, one entry per segment kotoba-whisper's chunking pipeline produced. That's the whole job. If you've used ComfyUI's list-processing conventions before (a node whose output is marked as a list feeds a for-each style downstream flow), this is exactly that pattern applied to speech segments.

The inputs and outputs that matter

• segments (required, KOTOBA_WHISPER_SEGMENTS) - the bundled segments object from either kotoba-whisper transcribe node in this pack that produces segments.
• segments (output, KOTOBA_WHISPER_SEGMENT, list) - the same data, exploded into individual items. Feed each one into SDT_KotobaWhisperSegmentProperty to pull out that segment's start, end, and text.

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 - this node is pure plumbing, nothing to download.

Common issues & troubleshooting

Fed in output from SDT_KotobaWhisperTranscribe (the generic, non-Short/Long node) and got an error. That node doesn't produce a segments output at all - it only returns flat text. This node needs segments to come from SDT_KotobaWhisperTranscribeShort or SDT_KotobaWhisperTranscribeLong specifically.

Not sure what to do with the exploded list. Wire each item into SDT_KotobaWhisperSegmentProperty to actually read its start, end, and text values - this node only splits the bundle apart, it doesn't expose those fields itself.

Empty or single-item list on a clip you expected multiple segments from. That reflects however kotoba-whisper's underlying pipeline chunked the audio, not something this node controls - a short or single-sentence clip may legitimately produce just one segment. If you expected finer granularity, that's a transcribe-node/loader-side question (chunk settings on SDT_KotobaWhisperLoaderLong), not something to debug here.

Wondering why this exists as a separate node instead of the transcribe node just returning a list directly. That's a ComfyUI graph-shape convention this pack follows consistently across its ASR families (the same List/Property split shows up for both the nemo-asr and faster-whisper sides): keep the raw model output as one bundled object, then let a dedicated node do the list explosion. It's more nodes to wire, but it keeps the transcribe nodes' output types simple and predictable.