This CHOP runs voice-activity-detection model. It takes an audio CHOP input and outputs a single channel Speaking (1.0 = speech, 0.0 = silence).
WHAT IT DOES
- Turns a live audio stream into a clean, frame-accurate is-someone-speaking signal.
- Real-time & lightweight. The model is ~2 MB and runs in well under a millisecond per window on the CPU.
- Speech, not just loudness. It distinguishes actual voice from music, fans, claps, and room noise, so it won't false-trigger on a loud non-speech sound the way a simple volume gate would.
- Stable output. Built-in hysteresis and hold time keep Speaking from flickering between words and breaths, and the streaming RNN state auto-resets after long silence so detection stays sharp over hours of input.
WHO IT'S FOR
- Anyone building interactive or live work in TouchDesigner who needs visuals, audio, or logic to react to when a person is talking.
- Live performers & VJs : gate reactive graphics, switch scenes.
- Installation artists : wake a piece when a visitor speaks, or branch a conversational/responsive experience on speech vs. silence.
- Broadcast & streaming operators : drive talking-head overlays, auto-frame the active speaker, or trigger lower-thirds and avatars only while someone is on mic.
- Generative & AI-pipeline builders : use it as a cheap front-end gate that only fires expensive downstream work (transcription, LLM calls, lip-sync) when actual speech is present.
