Load kotoba-whisper (Short-Form)

The dedicated short-form loader for kotoba-whisper, Kotoba Technologies' distilled Whisper model built specifically for Japanese. "Distilled" here means it's a smaller, faster student model trained to match a larger Whisper teacher's outputs on Japanese speech specifically, rather than leaning on general multilingual Whisper training - the upshot is meaningfully better Japanese accuracy and speed than reaching for a general-purpose Whisper model on the same audio.

How it works

"Short-Form" isn't decoration in the name - it matters mechanically. Whisper's encoder is built around a fixed ~30-second processing window; short-form loading means this node configures the model to run a clip through in a single pass rather than chunking it, which is both simpler and faster when your audio actually fits. Anything longer needs the chunked, sliding-window approach instead - that's what the separate long-form loader in this pack (and its matching SDT_KotobaWhisperTranscribeLong) handles.

The inputs and outputs that matter

• device - auto, cpu, or cuda. That's the entire configuration surface on this node - no chunking or batching options, since short-form doesn't need them. auto is the safe default; force cuda if you want a loud failure rather than an unexpectedly slow silent CPU fallback.
• model (output, KOTOBA_WHISPER_SHORT) - feeds directly into SDT_KotobaWhisperTranscribeShort, which is the only node in this pack that accepts this specific type.

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

Restart after. Model weights download from Hugging Face the first time you run this node - expect that fetch on first use, cached to disk afterward.

Common issues & troubleshooting

Clip is longer than roughly 30 seconds and results degrade, truncate, or only cover part of the audio. That's the short-form limitation showing up exactly as expected - this loader isn't built for long audio. Use the long-form loader instead, or split the clip first with SDT_TrimAudio/SDT_CutAudio if you specifically want to process it in short chunks yourself.

Feeding output into SDT_KotobaWhisperTranscribe (the generic one) and getting a type mismatch. The generic transcribe node expects KOTOBA_WHISPER, not KOTOBA_WHISPER_SHORT - they look similar but aren't interchangeable. This loader's output only pairs with SDT_KotobaWhisperTranscribeShort.

First run is slow or seems to hang. That's the model download, not a stall - there's no progress bar surfaced in the node UI for this. Check your console/server logs the first time you run it rather than assuming something's broken.

This is genuinely obscure territory - kotoba-whisper barely shows up outside its own Hugging Face model card, and this pack's wrapping of it has essentially no community footprint to search. If something behaves unexpectedly, the model card is your best reference, not a forum thread.