Load nemo-asr

This is the entry point for a third distinct Japanese ASR option in this pack, alongside kotoba-whisper and nue-asr. It loads the ReazonSpeech nemo-asr model, built by Reazon Human Interaction Lab on top of NVIDIA's NeMo toolkit - NeMo being NVIDIA's framework for training and running conversational-AI models, ASR included, and the "nemo-asr" naming here is literal: this is a NeMo-format acoustic model, not a Hugging Face transformers-style one like kotoba-whisper.

How it works

There's exactly one setting on this node - everything else about the model (architecture, tokenizer, training data) is fixed by what's packaged in the ReazonSpeech checkpoint itself. What makes this pairing worth knowing about specifically: SDT_NemoAsrTranscribe, the node this loader feeds, is the richest transcribe node in the whole pack - a single run gives you flat text, subword-level tokens with timestamps, and segment-level timestamps all at once, which is a direct reflection of what NeMo's ASR pipeline is built to expose natively (it tracks subword/token timing internally as part of decoding, not as an add-on).

The inputs and outputs that matter

• device - auto, cpu, or cuda. That's the entire input surface. auto picks sensibly; force cuda if you specifically want a loud failure rather than a silent, much slower CPU run.
• model (output, NEMO_ASR) - feeds directly into SDT_NemoAsrTranscribe, the only node in this pack that accepts it.

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. Model weights download from wherever the ReazonSpeech checkpoint is hosted on first use - expect a real fetch the first time, same as any pretrained model. NeMo's own Python package is a heavier install than a plain transformers-based model in general, so if setup here feels slower or pulls in more than you expected compared to the kotoba-whisper or faster-whisper loaders, that's the NeMo toolkit dependency, not something specific to this node.

Common issues & troubleshooting

First run is slow or looks stuck. That's the model download and NeMo's own initialization overhead, not a stall - NeMo models tend to have more setup cost on first load than a plain transformers pipeline. Check console/server logs before assuming it's broken.

Transcribing non-Japanese audio and getting poor results. The ReazonSpeech model is trained specifically on Japanese speech; feeding it other languages will give noticeably worse results than a general-purpose model. SDT_FasterWhisperTranscribe is the right choice for non-Japanese audio.

This is genuinely obscure territory. ReazonSpeech's nemo-asr model doesn't come up in general ComfyUI or speech-tooling discussion - it's a specialist Japanese ASR project, and this pack's wrapper around it has essentially no community footprint to search. The ReazonSpeech project itself is the better reference if something behaves unexpectedly.

Deciding between this, kotoba-whisper, and nue-asr for Japanese transcription. All three are legitimate Japanese-specific options in this pack and there's no single right answer without benchmarking your own audio - but this one is the node to reach for specifically if you want subword-level token timestamps alongside segment timestamps in one pass, which neither of the other two families exposes as directly.