faster-whisper Segment Property

Once SDT_FasterWhisperListSegments has fanned a transcript out into individual segments, this is the node that unpacks each one - pull the start time, end time, and text out of a single FASTER_WHISPER_SEGMENT so you can actually use them. This is the node you want if you're building a speech dataset: pair the text with the start/end window and you've got a labeled audio clip, one Whisper segment at a time.

How it works

Purely a property accessor - no processing, no model involved. It reads the three fields off a single segment object and exposes them as separate typed outputs so the rest of ComfyUI's graph (which works with individual FLOAT/STRING values, not opaque segment objects) can actually do something with them.

The inputs and outputs that matter

• segment (FASTER_WHISPER_SEGMENT) - a single segment, from SDT_FasterWhisperListSegments's list output. Since that node marks its output as a list, this property node runs once per segment automatically.
• start and end (both FLOAT) - the segment's timestamps in seconds. Wire these straight into SDT_TrimAudio's start_second/end_second to cut the exact audio slice that segment's transcript describes.
• text (STRING) - the transcribed text for just this segment. Pair it with the trimmed audio clip as your dataset label.

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 is pure property access, no ML runtime of its own.

Common issues & troubleshooting

Feeding start/end into SDT_TrimAudio and the resulting clip feels slightly off. Whisper-family timestamps are accurate to roughly the granularity of the model's internal windowing, not sample-exact - they're good enough for dataset work but not frame-perfect. If you need cleaner boundaries, pad the trim window a little on each side rather than trusting the raw timestamp exactly.

This node fires once per segment and you weren't expecting that. That's by design - it's meant to run per-item across the list SDT_FasterWhisperListSegments produces. If you only want the whole transcript as one string instead, skip both of these nodes and use SDT_FasterWhisperTextFromSegments directly off the transcribe node's output.

Text output has odd punctuation or capitalization. That's inherited straight from what faster-whisper produced for that segment - this node doesn't touch or clean the text at all, it just passes it through. If formatting matters for your dataset, that's a job for a text-processing node elsewhere in your graph, not this one.