Where knowing who is speaking changes everything.

The problemYour agent answers whoever makes noise: a TV in the background opens turns, a cough becomes a query, and the agent cuts users off mid-thought.

With VSIPSemantic endpointing waits out mid-thought pauses, noise events never reach your LLM, and speaker_verification means only the actual customer gets answered — from their very first sentence.

The problemPhone audio is 8kHz μ-law, callers interrupt constantly, and hold music or a second voice on speakerphone derails the bot.

With VSIPConnect Media Streams natively with audio_format=mulaw8k. Identity locks to the caller in their first sentence; barge-in works over the phone; overlapping speakers get separated channels.

The problemSpeech-to-speech stacks own their audio loop — you can't see who is speaking, and anyone near the microphone can steer the conversation.

With VSIPmode=sidecar runs listen-only beside the S2S agent: full identity, verification verdicts and noise events without touching the agent's audio. Drop responses when should_respond is false.

The problemSpeakerphones without echo cancellation hear their own TTS; busy environments mean five voices an hour at the same device.

With VSIPaec=server cancels the echo with a TTS reference channel, enabling strict identity gating: the enrolled user of the current session interrupts, everyone else is ignored until the session resets.

The problemMost stacks either ignore interruptions or stop for every sound. Users expect to say “actually, wait—” and be heard.

With VSIPbarge_in fires in ~250ms for the verified speaker, the interrupting utterance is a first-class turn (captured, verified, answerable), and your LLM is told its reply was cut off so it responds naturally.

The problemHealthcare and finance teams can't ship raw call audio to a third party that stores it.

With VSIPAudio is processed in memory and discarded — nothing is recorded. What persists is a 192-number voice embedding and billing metadata, isolated per tenant.