VVD Detect — the SDK

In any room with more than one microphone, the hard problem is not capturing audio — it is knowing who currently holds the floor, reliably enough to cut a camera on. VVD Detect listens across every microphone and answers that in real time. It also reads how the room feels, and it scales from a handful of microphones to hundreds.

What it tells your application

• The active speaker — one stable channel to put on screen, held sensibly through the natural pauses in speech rather than flickering on every breath.
• Who else is talking, and when two people talk at once — the cue for a wide shot.
• Silence, with a live meter and speaking-likelihood for every channel.
• The room’s mood — an optional read of the emotional tone (calm, tense, excited) that can drive shot width. It follows whoever is speaking on its own, from the audio you already send — nothing extra to wire up.

Built-in intelligence, no cloud

AI models are built into the library — one that tells speech from paper shuffles, chairs and air-conditioning, and one that reads the room’s emotional tone. Everything runs on the machine. Nothing is sent anywhere — no audio, no metadata, no phone-home — which is what lets it run on the firewalled and air-gapped networks broadcast facilities insist on.

Scales from one room to a whole rack

The AI runs across every channel in a single batched pass, so the same engine serves a small conference room or an installation with hundreds of microphones. For the largest channel counts it ships in GPU-accelerated builds — for NVIDIA (CUDA) and for Intel/AMD (DirectML) — drop-in with the standard CPU build: the same interface, you simply choose which file to ship.

What you get

A single native library with a simple, language-neutral C interface — active-speaker detection, AI voice-activity and room-mood sensing, with the models inside it. The standard CPU build is one self-contained file, no runtime to install. It comes with the same sample application written five times (C, C++, C#/.NET, Python and Rust), so you can start from whichever is closest to your codebase, plus a full integration guide.

How licensing works

Evaluate for free. The library runs unlicensed in a limited mode — the complete product, capped to a small number of channels — so you can build and test against your own audio before committing. Each licensed machine then gets a credential bound to that machine and verified entirely on the device, so it keeps working on an air-gapped network.