Pipecat VAD tuning cuts voice agent latency

Voice AI latency is often a "death by a thousand cuts" problem where VAD timeouts are the single biggest bottleneck for human-like conversation.

• –Reducing `stop_secs` from the default 0.5s to 0.15-0.2s is the most effective way to make a bot feel responsive, though it risks cutting off slower speakers.
• –Pipecat’s modular architecture allows developers to switch from local Silero VAD to server-side signals (e.g., via Sarvam STT) to eliminate local processing overhead and network jitter.
• –Sample rate mismatches, such as sending 48kHz audio to a 16kHz VAD, can introduce hidden resampling latency that compounds at each step of the pipeline.
• –Streaming LLM output is necessary but insufficient if the "first major blocker"—the decision that the user has finished talking—is delayed by conservative silence windows.