I'm here to help you integrate your own Voice Activity Detection (VAD) system instead of Silero in a LiveKit agent. Let's walk through the steps together: 1. First, create a new class that inherits from

agents.vad.VAD

(or the equivalent base VAD class in your SDK). 2. Next, implement the abstract

stream()

method to return an instance of your custom

VADStream

class, which should inherit from

agents.vad.VADStream

. 3. Within your

VADStream

, process the incoming

AudioFrame

objects, apply your VAD logic, and emit the appropriate

VADEvent

instances, such as

START_OF_SPEECH

and

END_OF_SPEECH

. 4. Finally, configure your agent or session to use your custom VAD instance instead of the default Silero VAD by passing it via the

vad

parameter in components like

AgentSession

or

VoiceAssistant

. As a helpful tip, the Silero VAD implementation is a great reference to see how audio frames are processed, inference is applied, and events are emitted. You can replace the ONNX model inference logic with your own algorithm while keeping the same event-based streaming interface. If you have any questions along the way, feel free to ask—I'm here to support you! Sources: AI Voice Assistant Quickstart | LiveKit Docs | Working with plugins | LiveKit Docs | agent/main.py | agents/src/voice/audio_recognition.ts