lifeiteng/omnivad-kit

OmniVAD-Kit is a toolkit for detecting when a person is speaking in an audio recording (Voice Activity Detection, or VAD) and for identifying other audio events like singing and music. These are common needs when building apps that transcribe speech, process meeting recordings, or need to know which parts of an audio file contain human voices versus silence or background noise. The toolkit provides three model types. The basic VAD model takes an audio file and returns timestamps of when speech occurs. The Stream-VAD model does the same thing in real time, processing audio frame by frame as it comes in (useful for live microphone input). The AED (Audio Event Detection) model can additionally distinguish between speech, singing, and music. All three underlying models are very compact, about 2.2 megabytes each, and run at roughly 200 times the speed of real time, meaning processing is nearly instant even on ordinary hardware. A key feature is that the toolkit runs in multiple environments without code changes. The Python package works on macOS, Linux, and Windows for server or desktop applications. A TypeScript/JavaScript package uses WebAssembly (a technology that lets compiled code run inside web browsers) so the same detection can run directly in a web page or Node.js server with no external service calls. A C API is also available for embedding into native apps on Android and other platforms. You would use this if you are building a transcription tool, a podcast editor, a meeting recorder, or any application that needs to locate speech within audio automatically.