Spectrogram

Same STFT machinery as SDT_MelSpectrogram, minus the perceptual mel warp. If you don't specifically need the mel scale - that matters for TTS/vocoder training, where models are built around it - this gives you the more literal representation: raw linear-frequency magnitude over time, no filterbank projection in between.

How it works

It's a short-time Fourier transform: chop the waveform into overlapping windows, FFT each one, and keep the resulting linear-frequency magnitude bins directly rather than passing them through a mel filterbank. The output type is SPEC, the same generic type SDT_MelSpectrogram produces - both feed into SDT_PlotSpectrogram interchangeably.

The inputs and outputs that matter

• audio - the clip to analyze, required.
• n_fft (default 400) - FFT window size, matching SDT_MelSpectrogram's default. Bigger trades time resolution for frequency resolution; the default is a well-worn value most people don't need to touch.
• win_length and hop_length (optional, both default -1) - leave at -1 for torchaudio's standard derived defaults (window length equal to n_fft, hop half of that). Set explicitly only if you need a specific frame rate to match something else in your pipeline.
• spec (output, SPEC) - the computed linear spectrogram.

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
pip install -r requirements.txt

No extra dependencies - pure signal processing on top of torchaudio, nothing to download.

Common issues & troubleshooting

Not sure whether you want this or SDT_MelSpectrogram. If you're feeding a TTS or vocoder pipeline, use SDT_MelSpectrogram - that's the representation those models are actually trained on. Use this node for general-purpose spectral analysis or visualization where mel warping isn't the goal.

Comparing two SPEC outputs and the frame counts don't match. hop_length controls that - mismatched hop lengths between two spectrograms (or between this node and SDT_MelSpectrogram) give you differently-shaped arrays even off identical audio. Set it explicitly and match it across both if you need aligned shapes.

Left win_length/hop_length at -1 and it looks fine. That's expected - torchaudio's derived defaults are sane for typical speech audio. Only override them if you need to match a specific frame rate elsewhere in your pipeline.

Feeding this SPEC into something that expects a mel spectrogram and getting bad results. A linear spectrogram and a mel spectrogram are both the pack's generic SPEC type, so ComfyUI's type system won't stop you from wiring one where the other is expected - but they carry genuinely different data (raw linear-frequency bins here versus perceptually mel-warped bins from SDT_MelSpectrogram). If a downstream model or node was trained on or expects mel features specifically, swapping in this node's output instead will run without erroring but produce nonsense. Know which one a given downstream node actually wants before wiring it in.