Load Silero VAD

Voice activity detection - figuring out which parts of a recording actually have someone talking versus silence, room noise, or dead air - is one of those unglamorous steps that makes everything downstream better. Feed a transcription model 10 minutes of mostly silence and you waste compute and invite hallucinated text in the gaps. Silero VAD is the model most people reach for to solve that: it's tiny, ONNX-based, and fast enough to run on CPU in real time. This node loads it.

How it works

Silero VAD (from the Silero team) is a small neural net trained specifically to classify short audio windows as speech or not-speech, exposed here as an ONNX Runtime session. This loader just instantiates it - it doesn't take an audio input and doesn't do any detection itself. That happens downstream in SDT_SileroVADApply, which takes the model this node outputs plus an actual clip.

The inputs and outputs that matter

• force_onnx_cpu (default false) - the only input. Silero VAD is cheap enough that it usually doesn't need a GPU at all; this flag forces the ONNX runtime session onto CPU even if a CUDA-capable onnxruntime is installed. Useful if you want to keep your GPU free for a heavier model running in the same graph, or if your onnxruntime-gpu install is flaky and you just want the model to load reliably.
• model (output, type SILERO_VAD) - wire it straight into SDT_SileroVADApply. It has no other consumer in this pack.

How to install it

ComfyUI Manager: search ComfyUI-speech-dataset-toolkit and install. Manual:

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

Then - and this is the part specific to this node - the README calls it out separately: pip install onnxruntime-gpu. It's not bundled in requirements.txt, so if you skip it, this loader will fail the moment you try to use it, even though the rest of the pack works fine. If you're CPU-only, install plain onnxruntime instead and set force_onnx_cpu to true.

Common issues & troubleshooting

Loader errors immediately with an import failure. Almost always the missing onnxruntime install. The pack's other nodes don't need it, so it's easy to git-clone, install requirements.txt, and assume you're done - then hit this the first time you actually reach for VAD.

GPU onnxruntime installed but still crashing/conflicting. onnxruntime-gpu is notoriously picky about matching CUDA/cuDNN versions with whatever else is in your environment (torch's own CUDA build included). If loading errors reference CUDA provider initialization rather than a missing module, set force_onnx_cpu to true - Silero VAD is small enough that CPU inference is genuinely fine for most audio lengths, so this isn't much of a downgrade.

Loads fine, does nothing useful on its own. That's expected - this node only loads the model. Actual detection, and the tunables that matter (threshold, minimum speech/silence duration), live on SDT_SileroVADApply.