Load nue-asr

The third and last Japanese ASR family in this pack, alongside kotoba-whisper and nemo-asr. This loads nue-asr, rinna Co., Ltd.'s speech recognition model - the one in this pack with the most distinct design goal of the three: rather than aiming for a strict phonetic-verbatim transcript the way Whisper-family models do, nue-asr pairs a speech encoder with a language-model decoder specifically to produce readable, natural Japanese text, closer to what a human transcriber would write than a literal word-for-word capture.

How it works

This node itself is a plain loader - it just configures how the model gets instantiated in memory, with two extra performance knobs beyond the usual device choice that the other loaders in this pack don't expose. The actual transcribe step happens in a separate nue-asr transcribe node elsewhere in the pack, not covered in this batch of articles, which takes this node's model output alongside an AUDIO clip.

The inputs and outputs that matter

• device - auto, cpu, or cuda. Same convention as every other loader in this pack: auto picks sensibly, force cuda if you want a loud failure instead of an unexpectedly slow silent CPU fallback.
• fp16 (default false) - run the model at half precision. Turning it on roughly halves memory footprint and speeds up GPU inference, at a small, usually negligible accuracy cost. Worth trying first if you're tight on VRAM before assuming you need a smaller setup.
• use_deepspeed (default false) - enables DeepSpeed-based inference optimization, a technique for speeding up transformer inference through kernel-level optimizations. This is the one setting on this node that adds real complexity: it's only useful if you already have a working DeepSpeed install in your environment, and turning it on without one is more likely to break your load than speed it up. Leave it off unless you specifically know you need it and have already set DeepSpeed up.
• model (output, NUE_ASR) - feeds into this pack's nue-asr transcribe node.

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 nue-asr checkpoint is hosted on first use. If you plan to try use_deepspeed, that's a separate, nontrivial install on top of this pack's baseline requirements - don't assume it's covered by requirements.txt without checking.

Common issues & troubleshooting

Turned on use_deepspeed and loading fails or errors oddly. The most likely cause: DeepSpeed isn't actually installed and working in your environment. This setting assumes it's already there - it doesn't install or configure DeepSpeed for you. Turn it back off unless you've specifically set DeepSpeed up and confirmed it works outside this node first.

Out of memory on GPU. Try fp16 before assuming you need more VRAM - it's the direct, low-effort lever on this loader for cutting memory footprint.

Transcript reads more like natural prose than a literal word-for-word capture, and that's throwing off downstream processing that expected verbatim text. That's not a bug - it's what nue-asr is specifically built to do, unlike a Whisper-family model. If your pipeline needs strict verbatim transcription (for precise timestamp alignment against every spoken word, for instance), SDT_NemoAsrTranscribe or a kotoba-whisper node is likely the better fit than nue-asr for that specific need.

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