VAD Voice Activity Detection
Detect when someone is talking versus silent in an audio stream — a critical step before speech recognition and in voice assistants.
Classic implementations: WebRTC VAD (Google, ~50 ms latency), Silero VAD (PyTorch model, more accurate). Modern: larger ML models with smarter VADs (Pyannote VAD segmentation, NeMo MarbleNet). Critical in voice-agent pipelines (LiveKit Agents, Pipecat, Vapi) to decide when the user has stopped talking ⇒ send to the LLM. "Endpointing" is non-trivial (people pause mid-sentence). Competitors on the voice front: ElevenLabs Conversational AI, OpenAI Realtime API integrate VAD internally.