Glossary

The words that appear in vendor decks, contracts and regulations, defined for the person signing the purchase order.

A

AI receptionistalso AI answering service, virtual receptionist, AI phone answering, AI front desk: A packaged voice agent sold to small businesses to answer the phone, take messages, answer common questions and book or move appointments, usually priced as a monthly subscription.
All-in cost per minutealso total cost per minute, fully loaded cost per minute, cost per connected minute, voice agent TCO, true cost per call: What one connected minute of agent conversation actually costs once platform fee, speech recognition, language model, text-to-speech, telephony, integration upkeep and human handling of transfers are all counted. Usually a multiple of the advertised per-minute price.

B

Bake-off (vendor evaluation)also vendor bake-off, proof of concept (PoC), pilot comparison, head-to-head evaluation, vendor trial: A side-by-side trial in which several vendors' agents face the same test material (your recordings, your script, your callers) and are scored on the same sheet. The alternative to choosing from vendor-run demos.
Barge-inalso interruption handling, interrupt, talk-over, cut-in, user interruption: The caller interrupting while the agent is speaking. A good agent stops within a fraction of a second, discards what it was about to say and takes the new input; a poor one talks over the caller or stops for a cough.

C

CCaaS (contact centre as a service)also cloud contact centre, contact center platform, CCaaS platform, contact centre software, UCaaS/CCaaS: The cloud platform that runs a contact centre: queues, routing, agent desktops, recording, workforce management and reporting. A voice agent either lives inside it as a feature, sits in front of it, or connects to it for transfers.
Code-switchingalso mixed-language speech, Hinglish, Taglish, Singlish, multilingual speech, language mixing: Switching between languages within a conversation or a single sentence, as in Hindi-English or Tagalog-English. Common in India, Southeast Asia and diaspora communities everywhere, and a hard case for speech recognition set to one language.
Containment ratealso automation rate, self-service rate, resolution rate, deflection rate, handled without agent: The share of calls the agent completes without a human, where 'completes' means the caller's goal was achieved and recorded in your system of record. Not the share of calls that simply ended without a transfer.
Conversational AIalso conversational agent, dialogue system, virtual agent, intelligent virtual assistant: The umbrella term for software that talks with people in natural language, across chat, messaging and voice. A voice agent is conversational AI on a phone line, with the timing and telephony problems that adds.

D

Do-not-call registryalso DNC list, Do Not Call Register, Telephone Preference Service (TPS), Do Not Disturb (DND), National Do Not Call Registry, opt-out list: A national list of phone numbers whose owners have opted out of marketing calls. Outbound voice-agent campaigns must be scrubbed against the relevant registry before dialling, and must also honour the business's own internal opt-out list.
DTMF (touch-tone)also touch-tone, keypad input, dual-tone multi-frequency, key press, press 1: The tones a phone keypad sends when a caller presses a key. Still the most reliable way to capture long numbers and unambiguous consent on a phone call, and a fallback when speech fails.

E

End-of-turn detectionalso turn detection, end-of-utterance detection, EOT detection, turn-end prediction, end-of-speech detection: The decision that the caller has finished speaking and the agent may reply. Too eager and the agent talks over people mid-thought; too cautious and every turn carries a long pause.
Endpointingalso endpoint detection, utterance end, silence threshold, endpointing threshold, final transcript timing: The speech-recognition mechanism that finalises a transcript after a stretch of silence. The endpointing threshold is often the single setting that most changes how quickly, and how often wrongly, an agent decides you have finished.
EU AI Act Article 50also AI Act transparency obligations, AI disclosure EU, Article 50 AI Act, Regulation (EU) 2024/1689 Article 50, AI Act chatbot disclosure: The transparency article of Regulation (EU) 2024/1689. People interacting with an AI system must be informed they are doing so unless it is obvious from context; synthetic audio must be marked as artificially generated. Transparency obligations apply from 2 August 2026, subject to pending amendments.

F

FDCPA and Regulation Falso Fair Debt Collection Practices Act, Regulation F, Reg F, 12 CFR 1006, seven-in-seven rule: The United States Fair Debt Collection Practices Act and the CFPB rule that implements it. Regulation F presumes harassment above seven call attempts per debt in seven days, or within seven days after a conversation, and restricts the times and places a collector may call.

G

Golden recordingalso golden set, reference recording, test set, ground-truth audio, reference transcript, evaluation corpus: A real call recording, consented and redacted where required, paired with a human-checked reference transcript. A set of them is the fixed yardstick for measuring speech accuracy and replaying the same caller behaviour against every vendor.

H

Hallucinationalso fabrication, confabulation, made-up answer, phantom transcript, false transcript: Confident output with no basis in the input. In a voice agent it happens twice over: the speech recogniser can produce words from silence or noise, and the language model can invent facts, prices or bookings. Both sound fluent.
HIPAA business associate agreement (BAA)also BAA, business associate agreement, HIPAA BAA, HIPAA compliance for voice agents, business associate: The contract a United States healthcare provider or insurer signs with any vendor that handles protected health information on its behalf. A voice agent that hears patient names and appointment reasons is such a vendor, as is every speech and language-model provider behind it.

I

IVR (interactive voice response)also phone menu, phone tree, auto attendant, speech IVR, directed-dialogue IVR: The menu-driven phone system that says 'press 1 for sales' or accepts a few fixed spoken keywords. The predecessor of the voice agent, and often still the front door it sits behind.

K

Knowledge base (RAG)also RAG, retrieval-augmented generation, knowledge source, grounding data, FAQ content, document retrieval: The documents, FAQs and policies the agent is allowed to answer from, and the retrieval step (retrieval-augmented generation) that pulls the relevant passage into the model's context before it replies.

L

Large language model (LLM)also LLM, language model, foundation model, the model, generative AI model: The model that reads the transcript, decides what the caller wants, chooses which tools to call and writes the reply. It sets the agent's reasoning quality, most of its latency and a large share of its per-minute cost.

P

PECR (UK)also Privacy and Electronic Communications Regulations, PECR Regulation 19, UK automated calling rules, ICO telephone marketing guidance, TPS: The United Kingdom's Privacy and Electronic Communications Regulations 2003. The ICO treats conversational AI voice calls as automated calls under Regulation 19, so direct marketing by a voice agent needs the recipient's specific prior consent, stricter than the rules for live human calls.
PSTNalso public switched telephone network, the phone network, landline network, telephone network, carrier network: The public switched telephone network: the global system of carriers and numbers that connects ordinary phone calls. Callers reach a voice agent through it, and its narrow-band audio and variable quality set a ceiling on what the agent can hear.

S

SIP trunkalso SIP trunking, SIP connection, VoIP trunk, BYOC (bring your own carrier), SIP forwarding: The internet-based connection (Session Initiation Protocol) that carries phone calls between a carrier or your phone system and the voice-agent platform. The usual way to connect an agent without moving your numbers.
Speaker diarizationalso diarization, speaker separation, speaker labelling, who-spoke-when, speaker attribution: Labelling a transcript with who spoke each segment (speaker 1, speaker 2). Matters for call analytics, quality review and multi-party calls; mostly irrelevant to a live two-party agent call, where the channels are already separate.
Speech-to-text (STT/ASR)also ASR, automatic speech recognition, speech recognition, transcription engine, STT: The component that turns the caller's speech into text for the language model. Its accuracy on your accents, languages and telephone audio caps everything the agent does afterwards.
Synthetic caller (agent-to-agent testing)also agent-to-agent testing, simulated caller, AI caller, automated conversation testing, persona testing, voice agent simulation: An automated caller, itself a voice agent, that phones the agent under test with scripted personas and goals, hundreds of times, so behaviour can be measured at scale. Scored by a rubric and a language-model judge, with human spot-checks because the judge is fallible.
System of recordalso SoR, source of truth, authoritative system, back-end system, core system: The application your organisation treats as the authoritative source for a fact: the practice schedule, the CRM, the policy administration system, the ticketing queue. A voice agent's outcome only counts when it is written there.

T

TCCCPR and DLT (India)also TCCCPR 2018, TRAI commercial communication rules, DLT registration, 140 series, 1600 series, DND scrubbing: India's Telecom Commercial Communications Customer Preference Regulations, administered by TRAI. Senders register on a Distributed Ledger Technology (DLT) platform, promotional calls use the 140 series and service calls the 1600 series, DND preferences are scrubbed, and promotional calls run 9 a.m. to 9 p.m. only.
TCPAalso Telephone Consumer Protection Act, TCPA consent, robocall rules, 47 CFR 64.1200, FCC AI voice ruling: The United States federal law restricting automated and prerecorded calls. The FCC confirmed in 2024 that AI-generated voices are 'artificial' voices under it, so outbound voice-agent calls need prior express consent and must follow calling-hour and identification rules.
Text-to-speech (TTS)also TTS, speech synthesis, synthetic voice, voice synthesis, neural voice: The component that turns the agent's text reply into a spoken voice. Buyers judge it on how quickly the first word arrives, how it pronounces names, numbers and addresses, and whether it sounds like a person on a telephone.
Time to first audio (TTFA)also TTFA, time to first byte of audio, first-audio latency, TTS latency: How long after the agent has something to say the caller hears the first sound. Usually quoted for the text-to-speech step alone; buyers should also ask for the end-to-end version measured from the end of caller speech.
Tool calling (function calling)also function calling, tool use, actions, API actions, skills, webhooks: The mechanism by which the language model asks the platform to run an action in your systems (check a slot, look up an account, book, transfer) and waits for the result before it speaks. It is how a voice agent does things rather than just talks.
Transfer ratealso escalation rate, handoff rate, agent transfer rate, fallback rate, human escalation: The share of calls the agent hands to a person. Some transfers are correct (emergencies, complex cases, callers who ask); others are failures. Only a reason on every transfer tells you which.
Turn-takingalso conversational turn-taking, floor management, duplex conversation, full-duplex handling: How the agent manages the back-and-forth of a conversation: when it starts talking, how long it waits, how it handles interruptions and overlaps, and how it recovers when both sides speak at once.

V

Voice activity detection (VAD)also VAD, speech detection, speech activity detection, speech presence detection: The detector that labels each slice of incoming audio as speech or not-speech. It gates when the agent starts listening, when it decides the caller has stopped, and whether an interruption counts.
Voice agentalso AI voice agent, voice AI agent, AI phone agent, conversational voice bot: Software that holds a spoken, two-way conversation with a caller over the phone or in an app, understands free speech, completes tasks in your systems and hands off to a person when it should.
Voice-to-voice latencyalso end-to-end latency, response latency, turn latency, mouth-to-ear latency, conversational latency: The time from the moment the caller stops speaking to the moment they hear the agent start to answer. The single number that most decides whether a voice agent feels like a conversation or like a phone menu.

W

Warm transferalso attended transfer, assisted transfer, handoff with context, supervised transfer, consultative transfer: A handoff in which the agent connects the caller to a person and passes along what has happened so far (transcript, summary, collected details), so the caller does not start again. A cold transfer drops the caller into a queue with nothing.
Word error rate (WER)also WER, transcription accuracy, recognition accuracy, character error rate (CER): The standard accuracy metric for speech recognition: substitutions plus deletions plus insertions, divided by the number of words actually spoken. Lower is better. Meaningless without knowing what audio it was measured on.
Write-backalso write integration, two-way integration, bidirectional sync, create/update actions, CRM write: The agent creating or changing a record in your system of record during the call: a booked appointment, an updated address, a logged payment promise. The acceptance test for any integration claim.