This category runs on vocabulary. Much of it is accurate; some of it is designed to sound more impressive than the thing it describes. This glossary covers only the terms that change a buying decision, and flags where the definition usually hides something.
The core pieces
AI voice agent. A system that holds a spoken conversation and can take actions — booking, recording, transferring. Distinct from a voice assistant mainly in emphasis: agents do things, assistants answer. The line is blurry in marketing and worth pinning down in a demo.
AI receptionist. A voice agent applied to inbound calls, positioned as a replacement for or supplement to a human front desk. "Receptionist" describes the role, not the technology.
ASR / STT (automatic speech recognition / speech-to-text). Turning audio into text. The two terms mean the same thing and are used interchangeably. This is the layer where most accuracy differences between languages and accents originate.
TTS (text-to-speech). Generating the spoken voice the caller hears. Modern TTS is genuinely good; if an assistant sounds robotic, it is usually a configuration or voice-choice problem rather than a technical limit.
LLM (large language model). The component that decides what to say next. This is where reasoning, tone and tool selection happen.
Pipeline. The chain ASR → LLM → TTS. Every stage adds delay, which is why latency is a system property rather than a single number.
The things you will actually notice
Latency. The gap between the caller finishing and the assistant starting. Text tolerates seconds of it invisibly. Voice does not — callers hear silence as failure. This is the single biggest reason a chatbot "ported to voice" disappoints.
Turn detection (endpointing). Deciding that the caller has finished speaking. Too eager and it interrupts people mid-thought; too patient and it feels sluggish. Good turn detection is invisible; bad turn detection is the most common complaint about voice systems.
Interruption handling (barge-in). What happens when the caller talks over the assistant. Real callers do this constantly. A system that stops cleanly and resumes is fundamentally better than one that talks through them.
Code-switching. A speaker mixing two languages in one sentence — extremely common among bilingual and newcomer callers ("I need an appointment for my hijo"). Handling it is much harder than handling either language alone, and most vendors do not claim it.
Hallucination. The model confidently producing something untrue. In voice this is worse than in text, because there is no draft state and the caller has already heard it.
The metrics — and how they mislead
Word error rate (WER). The standard measure of transcription accuracy — the proportion of words transcribed incorrectly. Lower is better, but it is not comparable across languages. Peer-reviewed audits have measured roughly 0.071 for well-supported English varieties and about 0.172 for speakers whose first language is tonal — around 2.4 times worse. Accent, age of language acquisition and data imbalance all shift it. Any single averaged WER conceals the variation that matters most.
Containment rate. The proportion of conversations the AI handled without a human. This is the most abused metric in the category. Ask one question: gross or net?
- Gross containment counts a conversation as contained if the AI handled it — including cases where the caller then called back, or gave up.
- Net containment subtracts callers who asked for a human immediately, escalated, or returned. Indications from published benchmarks suggest the gap between gross and net can be substantial.
If a vendor quotes containment without qualification, assume gross, and ask for net.
Handoff rate (escalation rate). How often a conversation reaches a human. This is arguably the most honest single figure available, because it shows where the system gives up. A high handoff rate in one language and a low one in another is your accuracy problem, surfacing.
Resolution rate. The share of conversations where the caller's need was met. Frequently conflated with containment; they are not the same. A call can be contained and unresolved.
Deflection. A call prevented from reaching a human. Useful to a vendor; potentially frustrating to a caller. When a public body reports deflected calls, it is describing people who were not offered a person at all.
The architecture terms
Knowledge base. The curated information the assistant answers from — policies, hours, services, FAQs. Quality here usually matters more than model choice.
RAG (retrieval-augmented generation). Looking things up in your knowledge base at answer time rather than relying on the model's memory. This is what keeps answers grounded in your actual business rather than plausible invention.
Function calling / tool use. The mechanism that lets the assistant act — check availability, write a record, send a message. Without it you have a system that can describe your business but not operate in it.
API and webhook. An API is how software asks another system to do something (check this calendar). A webhook is how a system notifies another that something happened (the call ended, here is the summary). "Integration" usually means a combination of the two.
SIP / PSTN / telephony. The plumbing that carries actual phone calls. PSTN is the traditional public telephone network; SIP is the protocol used to connect calls over the internet. You do not need to understand either, but if a vendor cannot explain how their system connects to a phone number, that is a warning.
DID (direct inward dialing). A phone number that routes to a specific destination. Relevant because most buyers keep their existing number — see whether AI voice works with your existing phone system.
Structured outputs. Forcing the system to return data in a fixed shape (name, phone, service, preferred time) rather than free text, so it can be written reliably into a CRM. Quietly one of the most important capabilities for a booking assistant.
Guardrails. Deliberate constraints on what the assistant may say or do — never quote a price, never give medical advice, always transfer for emergencies. Good guardrails are specific and testable.
Privacy and compliance
Call recording consent. Recording a call generally requires informing the caller. Rules differ by jurisdiction and by whether you are a party to the call.
PIPEDA / PHIPA / Law 25 / PIPA. The Canadian privacy statutes that apply to voice recordings and personal information — federal (PIPEDA), Ontario health (PHIPA), Quebec (Law 25) and British Columbia (PIPA). Our guide to AI voice assistants and Canadian privacy covers the practical obligations.
Retention. How long recordings and transcripts are kept. Ask for the number, not the policy summary.
Terms that should make you ask a second question
"Unlimited languages." Language coverage is not the same as equivalent quality. Supported and unsupported languages can differ by roughly a factor of two in error rate, and tonal languages perform worst.
"97% accuracy." Accuracy of what, measured how, on which speakers, in which language? A number without a method is a slogan.
"Seamless integration." Ask for a live demonstration of your specific action on your specific system.
"Fully autonomous." Almost always means "no human review", which is a liability rather than a feature in any setting where an error has consequences.
If a vendor cannot define their own metrics precisely, that is your answer — and it is more informative than any figure on the dashboard. For the wider picture, start with what is a voice AI assistant, or see what makes a good AI voice assistant for the seven things worth checking.