Voice AI glossary: conversation, telephony and outcome terms

Use this glossary to ask precise questions about a voice assistant. A term describes a concept; its presence in documentation does not prove that a particular product or configuration implements it.

Conversation and speech

Voice assistant: Software configured to respond through speech. Its permitted knowledge and actions depend on the application.

Cascade: A speech-recognition, language-model and speech-synthesis sequence. The stages can use separate providers.

Realtime speech model: A model that processes and generates audio without requiring the application to assemble the same separate text-based stages. See LiveKit's model overview for the architectural distinction.

STT / ASR: Speech-to-text or automatic speech recognition converts spoken input into text. Recognition can be wrong, especially for names, numbers or noisy input.

TTS: Text-to-speech produces audio from written text. Voice quality and recognition quality are separate properties.

VAD: Voice activity detection estimates whether speech is present. It is not, by itself, a complete understanding of whether a person has finished a thought.

Endpointing: Deciding that the caller's turn has ended. Silence, recognition signals and semantic information can contribute.

Interruption / barge-in: A caller begins speaking while the assistant is speaking, and the system decides whether and how to yield. A short acknowledgment is not always an intended interruption. LiveKit's turn guide explains these distinctions.

Response onset: The time at which the assistant's answer begins, measured against a defined reference event. State whether this means generated audio or playback heard by the caller.

Instructions, knowledge and actions

Prompt: Instructions and context supplied to a model. A prompt cannot replace authorization or make an unsupported integration work.

Context window: The input capacity a model can consider for an operation. Available history is not the same as durable memory across calls.

RAG: Retrieval-augmented generation supplies selected source material to help answer a question. Retrieval does not guarantee that the selected material is relevant, current or correctly used.

Hallucination: A plausible but unsupported or incorrect model output. Verify important answers and external actions against authoritative evidence.

Tool / action: A defined operation the assistant can request, such as availability lookup. The application must enforce permissions and validate arguments.

Structured extraction: Recording specified values from a conversation in a defined shape. A valid shape does not prove the extracted values are true.

Workflow: Explicit conversation stages and transitions. A selected runtime must support the nodes used by the workflow.

Idempotency: A mechanism that prevents repeated requests for the same operation from creating repeated effects, within the system's defined scope.

Reconciliation: Determining the actual result of an operation whose response was lost or uncertain before deciding whether to retry.

Telephone and browser connections

SIP: Session Initiation Protocol, used to establish, modify and terminate sessions. See RFC 3261.

SIP trunk: A configured connection between a phone system or voice service and a telephony provider. Account and route settings affect what it can carry.

PSTN: The public telephone network through which ordinary telephone numbers can be reached.

VoIP: Voice carried over Internet Protocol networks.

WebRTC: Browser-oriented realtime communication capabilities, standardized in the W3C WebRTC specification. A working browser session does not prove a telephone route.

DID / phone number: A telephone number routed to a destination. Recurring rental can continue even when an assistant is paused.

Caller ID: Presented calling information. It should not be treated as sufficient proof of identity for sensitive access.

Cold transfer: Moving a caller to another destination without a preceding private consultation.

Warm transfer: A transfer process that includes consultation before connecting the caller. Context delivery, rejection handling and recovery depend on the implementation.

Outcomes and accounting

Call completion: The session reached an end state. It does not establish a successful business action.

Verified outcome: The expected result is supported by evidence, such as a returned booking identifier or a correctly captured request.

Containment: Calls that remain within automation under a stated definition. It can include unresolved callers unless resolution is measured separately.

CSAT: Customer satisfaction measured from feedback. Automated sentiment or an evaluator score is not the customer's own rating.

Reservation: Funds or exposure set aside before usage. It is not the final charge.

Settlement: Applying the authoritative usage charge and releasing any unused reservation according to the billing rules.

BYO credentials: Customer-supplied provider credentials. They change provider access and billing responsibility, not automatically the platform's hosting or data-processing location.

For a practical comparison, use the voice AI testing guide. For privacy or regulatory questions, use the separate security review worksheet and the relevant primary guidance instead of treating an acronym as a certification.