Transcribe by nue-asr

SDT_NueAsrLoader gets nue-asr into memory; this is the node that actually runs it against a clip and hands you text back. The thing worth knowing going in: unlike the FasterWhisper and kotoba-whisper transcribe nodes elsewhere in this pack, there's no timestamped-segments output here. You put audio in, you get one plain string out. That's not a missing feature - it's downstream of what nue-asr is built to do.

How it works

nue-asr pairs a speech encoder with a language-model decoder specifically tuned to produce readable, natural Japanese text - closer to how a human transcriber would write something down than a strict phonetic-verbatim capture. Because the model's whole design goal is a clean, natural output rather than a timestamped decomposition of the audio, this node doesn't bother exposing segment boundaries the way the Whisper-family nodes do. It runs the model once over the whole clip and returns the resulting text.

The inputs and outputs that matter

• model (NUE_ASR) - from SDT_NueAsrLoader. No other node in the pack produces this type.
• audio - the clip to transcribe.
• text (output, STRING) - the finished transcript, ready to use as-is or pass into whatever text-processing you've got downstream.

That's the whole node - no tunable parameters beyond the two required wires.

How to install it

Via ComfyUI Manager: search ComfyUI-speech-dataset-toolkit, install, restart. Manually:

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   # match your CUDA version
pip install -r requirements.txt

Restart ComfyUI. This node itself adds nothing beyond the pack baseline - the model weights and any extra setup (like the loader's optional DeepSpeed path) live on SDT_NueAsrLoader, not here.

Common issues & troubleshooting

Errors before this node produces anything. Almost always a loader problem, not this one - check SDT_NueAsrLoader loaded successfully first, especially if you toggled its use_deepspeed option.

You expected timestamps or segments and got a single string. That's correct behavior, not a bug - nue-asr's whole point is readable prose output, not a timed breakdown. If you need per-segment timing, a kotoba-whisper or SDT_FasterWhisperTranscribe node is the right tool for that job, not this one.

Transcript reads more like natural prose than a literal word-for-word capture, and that's throwing off something expecting verbatim text. Also expected - nue-asr is deliberately optimized for readability over strict phonetic accuracy. If your pipeline needs a literal, word-for-word transcript for precise alignment work, a Whisper-family node in this pack will serve you better.

This is genuinely obscure territory. A search for community discussion around nue-asr turns up essentially nothing - it doesn't come up in general ComfyUI or speech-tooling threads the way Whisper does. If something behaves unexpectedly, the nue-asr project itself is a far better reference than searching for help online.