Make Silence Audio

Sometimes you don't need to process audio - you need to manufacture some. A fixed block of silence is genuinely useful when you're assembling a dataset: padding a clip out to a target length, inserting a clean gap between concatenated utterances, or building a negative example for a VAD/ASR pipeline that should output nothing. This node generates exactly that - a silent clip with the duration, sample rate, and channel layout you specify.

How it works

No audio input at all - it synthesizes a zero-amplitude waveform from scratch using torchaudio's tensor primitives, at whatever length and sample rate you ask for, and hands it back as a normal AUDIO object. Because it's just zeros, it composes cleanly with everything else in the pack: concatenate it onto real clips, trim it, resample it, whatever you need.

The inputs and outputs that matter

• second (default 0, step 0.001) - duration of the silence in seconds. The 0.001 step means you can dial in silence down to the millisecond, useful when you're padding to match an exact frame count elsewhere in a pipeline.
• sample_rate (default 16000) - match this to whatever the rest of your pipeline expects. 16kHz is the standard rate most speech models (Whisper included) are trained on; if you're feeding this silence into a music-oriented chain, you'll want something higher.
• channel - monoral or stereo. Match it to whatever you're padding or concatenating with; mismatched channel counts are a common source of shape errors when two AUDIO tensors meet downstream.
• audio (output) - a silent clip, ready to feed anywhere an AUDIO input is expected.

How to install it

ComfyUI Manager: search ComfyUI-speech-dataset-toolkit, install, restart. 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

This node has no extra dependencies beyond torchaudio, which the whole pack needs anyway.

Common issues & troubleshooting

Concatenating with real audio produces a click or glitch at the boundary. Check that sample_rate here matches the real clip's sample rate exactly - a mismatch will either error or, if your downstream node silently resamples, introduce artifacts right at the seam. SDT_AudioProperty is the quickest way to check a clip's actual sample rate before you set this.

Silence isn't actually silent in your saved file. Unlikely given it's a literal zero tensor, but if you're piping it through a lossy codec (mp3, vorbis) via SDT_SaveAudio, be aware lossy encoders can introduce a faint noise floor even on true silence - that's a codec property, not this node misbehaving.

Using it for VAD/ASR negative testing and getting unexpected results. Some models (and some VAD configs) treat true digital silence differently from real-world "quiet" recordings, which always carry some room noise. If you're testing a pipeline's behavior on quiet-but-not-silent input, this node's output is a poor stand-in - you'd want a real low-volume recording instead.