Transcribe by kotoba-whisper (Long-Form)

The long-form partner to SDT_KotobaWhisperLoaderLong: this is the node that actually transcribes once you've loaded kotoba-whisper's chunked pipeline for audio running past Whisper's native ~30-second window. If your Japanese clips are full recordings rather than short utterances, this - not the short-form transcribe node - is the one you want.

How it works

Whisper's encoder is built around a fixed-length window, so anything longer has to be split into overlapping chunks, transcribed piece by piece, and stitched back together - the chunking configuration (chunk_length_s, batch_size) lives on SDT_KotobaWhisperLoaderLong, and this node just runs that already-configured pipeline against your audio and returns both a flat transcript and the underlying timestamped segments the chunking process produced along the way.

The inputs and outputs that matter

• model (type KOTOBA_WHISPER_LONG) - required, must come from SDT_KotobaWhisperLoaderLong specifically; it won't accept a model from the short-form or generic loaders.
• audio - required, the clip to transcribe.
• No optional inputs - notably, unlike SDT_KotobaWhisperTranscribe and SDT_KotobaWhisperTranscribeShort, there's no prompt field here. If you're used to context-priming Whisper with an initial_prompt-style hint on the other transcribe nodes in this pack, it's simply not exposed on this one.
• text (output, STRING) - the flattened, full transcript across the whole clip.
• segments (output, KOTOBA_WHISPER_SEGMENTS) - the timestamped breakdown from the chunking pipeline. Feed this into SDT_KotobaWhisperListSegments to explode it into individual segments, then SDT_KotobaWhisperSegmentProperty per segment for start/end/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

Model weights download from Hugging Face on first use, through SDT_KotobaWhisperLoaderLong - this transcribe node itself doesn't add any dependency weight of its own.

Common issues & troubleshooting

Looking for a prompt input and it isn't there. That's not a bug - the long-form pipeline in this pack simply doesn't expose one, unlike the short-form and generic transcribe nodes. There's no workaround within this node; if prompt-priming matters for your use case, it isn't available on the long-form path.

Words repeated or dropped at what look like regular intervals. A classic symptom of chunk-boundary stitching in any long-form ASR pipeline. If it's bad enough to matter, go back to SDT_KotobaWhisperLoaderLong and try a larger chunk_length_s so there are fewer seams for the pipeline to reconcile.

Out of memory on a long file. Also a loader-side setting - drop batch_size on SDT_KotobaWhisperLoaderLong first, since it controls how many chunks run concurrently and is the more direct memory lever than chunk length.

This is genuinely obscure territory - kotoba-whisper's long-form path barely shows up outside its own Hugging Face model card, and this specific node has essentially zero community footprint. Treat the model card and this pack's source as your references rather than expecting a forum thread to exist.