Transcribe by kotoba-whisper

This is the transcribe node that pairs with the generic SDT_KotobaWhisperLoader - the one where you pick short-vs-long form on the loader itself rather than picking a dedicated loader node. What sets it apart from the pack's other two kotoba-whisper transcribe nodes (SDT_KotobaWhisperTranscribeShort and SDT_KotobaWhisperTranscribeLong) is what it doesn't give you: just flat transcript text, no segment timing at all.

How it works

It runs kotoba-whisper - Kotoba Technologies' distilled Whisper model trained specifically for Japanese - over the audio using whichever mode (short or long) you configured on the loader, and returns the resulting transcript as one string. There's no chunk-boundary stitching detail exposed to you here, no timestamped breakdown - the pipeline handles all of that internally and you just get the words out the other end. That makes this the node to reach for when you genuinely just want the text and nothing else, rather than building a dataset that needs per-sentence timing.

The inputs and outputs that matter

• model (type KOTOBA_WHISPER) - required, must come from SDT_KotobaWhisperLoader specifically. The dedicated Short/Long loaders produce different types (KOTOBA_WHISPER_SHORT/KOTOBA_WHISPER_LONG) that this node won't accept.
• audio - required, the clip to transcribe.
• prompt (optional, blank by default) - Whisper's context-priming trick: text you supply here biases the model's vocabulary and formatting toward it, useful for nudging proper nouns, technical terms, or a particular punctuation convention without fine-tuning anything.
• text (output, STRING) - the flat transcript, and the only output. If you need timestamps, this is the wrong node - see below.

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 SDT_KotobaWhisperLoader - this node itself adds nothing beyond that.

Common issues & troubleshooting

You need segment timing and this node doesn't have it. That's by design, not a missing feature - this is the flat-text-only variant. Switch to SDT_KotobaWhisperTranscribeShort or SDT_KotobaWhisperTranscribeLong (matched with their respective dedicated loaders) if you need the segments output alongside text.

Type mismatch feeding in a model from a different kotoba-whisper loader. The model input here specifically wants KOTOBA_WHISPER from SDT_KotobaWhisperLoader - the dedicated Short/Long loaders in this pack produce different, incompatible types.

Transcribing non-Japanese audio and getting poor results. kotoba-whisper is trained and distilled specifically for Japanese; feeding it other languages gives noticeably worse output than a general Whisper model would. Use SDT_FasterWhisperTranscribe instead if your audio isn't Japanese.

Long clip fed in with form_length set to short on the loader. Results will degrade or truncate - that setting lives on SDT_KotobaWhisperLoader, not here, so if this node's output looks wrong on longer audio, go back and check the loader's form_length rather than assuming this node mishandled it.