GLOSSARY

Voice AI glossary

I'm shopping for an AI phone system and the vendor pages are full of jargon — what do all these voice-AI terms actually mean?

Voice-AI phone systems rest on a few core ideas: speech-to-text turns a caller's words into text, an LLM decides what to say, and text-to-speech speaks the reply back — fast enough to feel human. This glossary defines the 18 terms you meet on vendor pages in plain English, so you can compare tools like Teeoff without a technical background.

UPDATED 2026-07

Speech-to-text (STT)

Speech-to-text is the software that converts a caller's spoken words into written text in real time, so the rest of the system can understand what was said. On a live phone call it has to keep up with natural speech — pauses, accents and background noise — without falling behind. Its accuracy is measured by word error rate. Related: ASR vs STT, word error rate (WER).

Text-to-speech (TTS)

Text-to-speech does the reverse of STT: it turns the words your AI agent decides to say into a natural-sounding spoken voice the caller hears. Good text-to-speech sounds human rather than robotic, and starts speaking quickly so the conversation doesn't feel laggy. How fast that first sound arrives is part of latency. Related: latency / time-to-first-token, TTS voice cloning.

Word error rate (WER)

Word error rate is the standard score for how accurate speech-to-text is: the percentage of words it gets wrong — misheard, missed or added — compared with what was actually said. Lower is better, so a 5% word error rate means 95 of every 100 words are correct. Accents, noise and phone-line quality push it up. Related: speech-to-text (STT), code-switching.

Latency / time-to-first-token

Latency is the delay between a caller finishing their sentence and the AI starting to respond. Measured end-to-end, it should feel like a natural human pause — long delays make callers talk over the agent or hang up. 'Time-to-first-token' is the slice of that delay before the AI produces its very first word. Related: barge-in, text-to-speech (TTS).

Barge-in

Barge-in lets a caller interrupt the AI mid-sentence — starting to talk while it is still speaking — and have the agent stop and listen, just as a person would. Without it, the agent talks over people and feels robotic. It depends on the system detecting incoming speech almost instantly. Related: voice activity detection (VAD), endpointing.

Endpointing

Endpointing is how the system decides a caller has finished speaking so the AI can take its turn. Cut it too early and you interrupt people mid-thought; cut it too late and the agent feels slow to answer. Good endpointing tells a natural pause apart from the real end of a sentence. Related: voice activity detection (VAD), latency / time-to-first-token.

Diarization

Diarization is the system working out who spoke when — separating and labelling different voices on a call, such as the caller versus your agent, or two people sharing one phone. It is what makes a transcript readable ('Caller:' versus 'Agent:') and is useful for call records and quality review. Related: speech-to-text (STT), PII redaction.

Voice activity detection (VAD)

Voice activity detection is the lightweight step that decides, moment to moment, whether someone is actually speaking or the line is just noise or silence. It is the trigger behind natural turn-taking: knowing when to listen, when a caller starts talking, and when to allow an interruption. Related: barge-in, endpointing.

Concurrency

Concurrency is how many phone calls your AI agent can handle at the same time. A human receptionist takes one call at a time; an AI can answer many at once, so nobody hits a busy signal during a rush. If a vendor caps concurrency, callers can still be turned away at peak. Related: SIP / telephony, IVR.

Warm transfer

A warm transfer is when the AI hands a call to a human and passes along context first — who is calling, why, and what has already been said — so the caller does not have to repeat themselves. A 'cold' transfer just dumps the call through with no background. Warm transfers make escalations feel seamless. Related: IVR, concurrency.

IVR

IVR (interactive voice response) is the old 'press 1 for sales, press 2 for support' phone menu. It routes callers using keypad presses or rigid voice commands. Modern AI voice agents replace rigid IVR trees with natural conversation — the caller just says what they need in their own words. Related: warm transfer, SIP / telephony.

LLM

An LLM (large language model) is the "brain" that decides what your AI agent says. It reads the transcribed caller question, works out the intent, and generates a helpful reply in plain language — handling follow-ups and context rather than following a fixed script. It is the same class of technology behind ChatGPT. Related: RAG, speech-to-text (STT).

RAG

RAG (retrieval-augmented generation) lets the LLM pull answers from your specific business information — hours, prices, policies, FAQs — instead of guessing or making things up. The agent looks up the relevant facts first, then uses them to answer accurately. It is how a voice agent knows your details, not just general knowledge. Related: LLM, PII redaction.

ASR vs STT

ASR (automatic speech recognition) and STT (speech-to-text) mean essentially the same thing — software that converts spoken words into text. 'ASR' is the older, more technical term; 'STT' is more common in product marketing. If a vendor lists one or the other, they are describing the same capability. Related: speech-to-text (STT), word error rate (WER).

Code-switching

Code-switching is when a caller mixes languages within one conversation — or even one sentence — such as switching between English and Spanish mid-thought. Many communities do this naturally. A voice agent that handles code-switching can follow along without forcing the caller to pick a single language. Related: speech-to-text (STT), text-to-speech (TTS).

TTS voice cloning

Voice cloning creates a custom text-to-speech voice that matches a specific person or a brand style, built from a short recording. A business might use it so its AI agent sounds consistent with its brand. Reputable providers require consent and safeguards to stop the feature being misused. Related: text-to-speech (TTS), PII redaction.

PII redaction

PII redaction automatically finds and hides personally identifiable information — names, card numbers, addresses, health details — in call recordings and transcripts. It keeps sensitive customer data out of stored logs, which matters for privacy and for rules such as HIPAA or PCI. Related: RAG, diarization.

SIP / telephony

SIP (session initiation protocol) is the standard plumbing that carries phone calls over the internet, connecting your business phone number to the AI agent. 'Telephony' is the broader term for that phone infrastructure. It is what lets an AI answer your existing number without new hardware. Related: concurrency, IVR.

Questions, answered

The AI phone vendors keep talking about "latency" — why should I care?

Latency is the pause between a caller finishing their sentence and the AI answering. If it is too long, callers talk over the agent or assume the line dropped and hang up; if it feels like a natural human pause, the conversation flows. It is one of the most noticeable quality differences between voice agents, which is why Teeoff optimises for a fast, human-feeling response.

A vendor page lists both STT and TTS as features — do I need both, and what do they actually do?

They are the two halves of a spoken conversation. Speech-to-text (STT) listens and converts the caller's words into text the system can understand. Text-to-speech (TTS) does the reverse — it turns the agent's reply into a natural-sounding voice the caller hears. A voice agent like Teeoff runs both, plus an LLM in between to decide what to say.

My customers switch between English and Spanish on a call — can AI handle that?

Yes, if the system supports code-switching — following a caller who mixes languages within one conversation or even one sentence, without forcing them to pick one. Not every voice agent does this well. Teeoff is built to be multilingual and to handle callers who move between languages naturally, which matters for businesses serving diverse communities.

Is my customers' data safe if an AI records and transcribes calls?

It depends on the provider's safeguards. Look for PII redaction, which automatically hides sensitive details — names, card numbers, health information — from stored recordings and transcripts, and for compliance with rules like HIPAA or PCI where they apply. These are the guardrails that keep call data from becoming a liability; ask any vendor, including Teeoff, how they handle them.