Load SpeechMOS

This loads SpeechMOS, UTokyo-SaruLab's model for predicting a Mean Opinion Score - the roughly 1-to-5 naturalness rating a human listener would typically give a clip in a formal listening test - without needing an actual human listening test. If you're iterating on a TTS or voice-conversion model and want a repeatable automatic quality signal instead of running listening panels every time you change a hyperparameter, this is the loader that gets you there.

How it works

This node itself just instantiates the model - device selection only, nothing else to configure. The actual scoring happens downstream in SDT_SpeechMOSScore, which takes this node's output alongside a clip.

MOS predictors like this one work by training a neural net directly on human listening-test ratings - thousands of clips, each with an actual average score from human raters - until the model learns to predict what a panel of listeners would probably say about a clip it's never heard. That's why it needs no reference audio to compare against: it's not measuring similarity to anything, it's predicting a rating. This kind of automatic MOS estimation is what let TTS and voice-conversion research scale past the point where a human listening panel for every single model checkpoint was practical.

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'd rather get a loud failure than an unexpectedly slow silent CPU fallback.
• model (output, SPEECH_MOS) - feeds into SDT_SpeechMOSScore. No other node consumes this type.

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 on first use, same pattern as the pack's other pretrained-model loaders.

Common issues & troubleshooting

Loads fine, does nothing useful on its own. Expected - this node only loads the model. Pair it with SDT_SpeechMOSScore to actually get a score out.

Forced cuda and it failed instead of falling back. By design, same reasoning as every other loader here: better to fail loudly and know immediately than to silently fall back to a much slower CPU run.

This is genuinely niche territory. A search for community discussion around SpeechMOS turns up essentially nothing relevant - unquoted searches mostly surface unrelated threads that happen to contain the word "speechless." It doesn't come up in general ComfyUI or TTS-tooling discussion the way mainstream evaluation tools do. If something behaves oddly, the SpeechMOS project itself is a far better reference than searching for help online.