Transcribe by kotoba-whisper (Short-Form)

If you're transcribing Japanese and faster-whisper isn't giving you the quality you want, this is the node built for that specifically. It runs kotoba-whisper, Kotoba Technologies' distilled Whisper model trained specifically for Japanese - distillation here means a smaller, faster student model trained to match a larger Whisper teacher's outputs on Japanese speech, rather than relying on Whisper's general multilingual training alone.

How it works

"Short-Form" in the name isn't decoration - it matters. Whisper's encoder natively processes fixed ~30-second windows; short-form transcription means the clip fits in one pass, no chunking needed. That's what this node does. Longer audio needs the chunking/sliding-window approach instead, which is what the separate SDT_KotobaWhisperLoaderLong (long-form) path in this pack is for - that loader isn't covered here, but it's worth knowing it exists if your clips run past 30 seconds, since this transcribe node pairs with a matching short-form loader elsewhere in the pack rather than the long-form one.

The inputs and outputs that matter

• model (type KOTOBA_WHISPER_SHORT) and audio - required. Note the model type is specifically the short-form variant; it won't accept a long-form loaded model.
• prompt (optional, blank by default) - the same context-priming idea as Whisper's initial_prompt: text you supply here biases the model's output vocabulary and style, useful for nudging transcription of proper nouns or a particular formatting convention without retraining anything.
• text (output, STRING) - the flat transcript, ready to use directly.
• segments (output, KOTOBA_WHISPER_SEGMENTS) - the timestamped breakdown, for anything downstream that needs per-line timing rather than just the words.

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 come from Hugging Face on first load, through whichever short-form loader you're pairing this with - expect a download the first time, same as any pretrained checkpoint.

Common issues & troubleshooting

This is a low-traffic corner of a low-traffic pack - kotoba-whisper doesn't come up in general ComfyUI community discussion the way faster-whisper does, so if you hit an edge case here, don't expect much prior art to search for. Read the kotoba-whisper model card directly if something looks off; it's the primary source for this model's actual behavior.

Transcribing something that isn't Japanese. This model is trained and distilled specifically for Japanese - feeding it other languages will give you noticeably worse results than a general Whisper model would. If your audio isn't Japanese, SDT_FasterWhisperTranscribe is the right node instead.

Clip is longer than ~30 seconds and results degrade or truncate. That's the short-form limitation biting - this node isn't built for long audio. Split the clip first (SDT_SplitAudio/SDT_TrimAudio) or use the long-form loader/transcribe path instead.