The words, without the marketing.

Voice agent

A voice agent is software you talk to out loud that hears you, works out what you are asking, decides what to do about it, and answers in speech.

Voice widget

A voice widget is the embeddable front end of a website voice agent: the script tag that puts it on the page, the isolated interface a visitor taps, and the microphone and audio-transport layers behind it.

Website voice agent

A website voice agent is a voice agent that runs on a web page rather than a phone line, so a visitor can ask a question out loud and hear the site answer it from the information that business has published.

Automatic speech recognition

Automatic speech recognition (ASR), also called speech to text, is the technology that turns spoken audio into written text, and it is the first real step in almost every voice agent.

Barge-in

Barge-in is a voice agent stopping its own speech the instant a person starts talking over it, which is what makes the exchange feel like a conversation rather than a recording you have to sit through.

Endpointing

Endpointing is how a voice agent decides that the person has finished speaking and its own turn has begun, normally by timing the silence against a threshold set between 150 and 300 milliseconds.

Speech-to-speech models

A speech-to-speech model is a single model that consumes audio and emits audio directly, with no transcription step in the middle, which buys latency and preserved tone at the cost of the text record a cascaded pipeline produces as a by-product.

Text to speech

Text to speech (TTS), also called speech synthesis, turns written text into spoken audio, and it is how a voice agent gets its voice.

Turn detection

Turn detection is working out whose turn it is to speak, and semantic turn detection does it by judging whether the sentence sounds finished rather than by timing the silence after it.

Voice activity detection

Voice activity detection (VAD) is the component at the very front of a voice agent that decides, frame by frame, whether the incoming audio contains speech or only background noise.

Voice cloning

Voice cloning builds a synthetic copy of one named person's voice out of a recording of them talking, and that copy can then be made to say things the person never said.

Mean opinion score

Mean opinion score (MOS) is a 1 to 5 quality rating for synthesized or transmitted audio, produced by asking human listeners to score samples and averaging their answers, where 5 is excellent.

Time to first audio

Time to first audio (TTFA) is the end-to-end wall-clock delay from the moment a person stops speaking to the moment the reply becomes audible, and it decides more than any other number whether a voice agent feels alive.

Word error rate

Word error rate (WER) is the standard measure of transcription accuracy: the words a speech recognition system got wrong, counted as a percentage of the words that were actually said, where lower is better.

Chunking

Chunking is splitting documents into smaller passages before they are indexed for retrieval, and where those splits fall largely decides whether the right answer can be found at all.

Context window

The context window is how much text a model can consider at once, counting both what it has been given and what it produces, and everything outside the window does not exist as far as the model is concerned.

Conversation memory

Session state is what a voice agent remembers within a single conversation and conversation memory is what it remembers across conversations, and the two are different problems carrying different risks.

Embeddings

An embedding is a list of numbers representing the meaning of a piece of text, arranged so that two passages about the same thing sit close together even when they share no words, which is what lets a system search by meaning instead of by keyword.

Function calling

Function calling is how a language model acts rather than only speaks: it emits a structured request for an action, and the application around it checks that request, runs it against a real system, and feeds the result back into the conversation.

Grounding

Grounding is the discipline of tying what an AI says to a verifiable source, so the answer can be traced to real material rather than to whatever the model absorbed during training.

Guardrails

Guardrails are the limits that keep a voice agent from saying or doing what it should not, whatever the visitor asked and whatever the system retrieved, and the thing that makes them guardrails is that they hold outside the model instead of being asked of it.

Hallucination

A hallucination is when an AI states something false with the same confidence it states something true, which is not a lie, because there is no intent, and not a bug, because the system is working as designed.

Human handoff

Human handoff is the act of moving a live conversation from an automated agent to a person, and escalation is the judgement that the move is needed, and together they are what a customer runs into whenever the automation could not finish the job.

Reranking

Reranking is a second pass that reorders retrieved results by relevance using a slower and more accurate model than the initial search, and it is what fixes the near-misses fast retrieval gets wrong.

Retrieval-augmented generation

Retrieval-augmented generation, or RAG, is a technique that makes an AI look the answer up in specific documents you have given it before it replies, instead of answering from whatever it absorbed during training.

Vector database

A vector database holds embeddings and returns the nearest ones to a query fast enough that nobody waits, which makes it the search index that meaning-based retrieval runs on.

Audio codecs

An audio codec is the scheme that squeezes sound into a form a network can carry, and whichever one a voice system runs on settles how the audio sounds, what it costs in bandwidth, and how much of it a recognizer can actually transcribe.

Concurrency

Concurrency counts the conversations a voice system can run at once, which makes it a question about capacity rather than about volume, and it governs what happens on the busiest day instead of the ordinary one.

DTMF

DTMF, dual-tone multi-frequency, is the touch-tone scheme that turns every keypad press into its own pair of tones, and it remains the right way to collect the card numbers, PINs and reference codes that speech recognition reliably mangles.

PSTN

The PSTN, or Public Switched Telephone Network, is the worldwide arrangement of carriers, numbering plans and interconnect deals that lets any phone number dial any other, and the narrowband audio it carries holds measurably less information than a browser microphone captures.

SIP and SIP trunking

SIP, the Session Initiation Protocol, is what phone systems use to open, manage and close voice calls across internet networks, and SIP trunking is the service you buy to reach the outside telephone network through an internet connection rather than through physical lines into the building.

STIR/SHAKEN and spam-likely labeling

STIR/SHAKEN is a framework that lets carriers cryptographically verify a call really originates from the number it claims, while spam-likely labeling is a separate and proprietary scoring layer that decides whether a warning appears on the recipient screen.

WebRTC

WebRTC is a set of open browser standards for capturing, processing and transporting live audio and video in real time, and it is what lets a website hold a spoken conversation with no plugin, no app and no phone call.

Call deflection

Call deflection is the practice of moving a customer contact out of an expensive channel, usually the phone, into a cheaper one such as self-service, and it counts the contact that never arrived rather than the problem that got solved.

Containment rate

Containment rate is the share of conversations an automated system finishes on its own without handing the person to a human, calculated as conversations resolved without human involvement divided by total conversations handled.

Answer engine optimization

Answer engine optimization is the practice of writing and structuring content so an AI answer engine picks it up and cites it, which means competing for a mention inside someone else's answer rather than for a ranking position on a results page.

Zero-click search

A zero-click search is a search where the person gets their answer on the results page itself and never visits a website, so the search succeeds while the site that had the answer receives nothing.

AI disclosure and the EU AI Act

AI disclosure means telling people they are interacting with an AI rather than a human, and the EU AI Act turns that into a legal obligation in defined circumstances when its Article 50 transparency provisions become enforceable on 2 August 2026.

CCPA and CPRA

The CCPA, as amended by the CPRA, gives California residents rights to know about, delete, correct and stop the sharing of the personal information a business holds, and it asks whether that business disclosed and can service requests rather than whether it had a basis to collect at all.

Consent and call recording law

Consent in a voice deployment is three separate permissions, one to record the conversation, one to use what it produced and one to contact the person at all, and holding any of them does not grant the others.

DPDP Act

The DPDP Act is the Indian data protection law, built around consent as the principal basis for processing, which makes it narrower than GDPR in exactly the place most international deployments assume it will be looser.

GDPR

The GDPR is the European Union data protection regulation, and it sets the rules for collecting and using personal data about people in the EU no matter which country the organization holding that data is registered in.

HIPAA

HIPAA is the United States law governing the privacy and security of health information, and any voice AI vendor that handles patient information on behalf of a healthcare provider becomes a business associate directly bound by it.

PCI DSS

PCI DSS is the contractual security standard the payment card industry imposes on anyone who stores, processes or transmits cardholder data, and for a voice agent the compliant approach is almost always to make sure it never touches card details at all.

PII redaction and data residency

PII redaction removes personally identifiable information from transcripts and recordings, while data residency governs which country that data is stored and processed in, and both are standard enterprise requirements with specific consequences for a voice pipeline.

SOC 2

SOC 2 is an independent audit report, issued under the AICPA Trust Services Criteria, giving an auditor's opinion on whether the controls an organization has described are suitably designed and, in a Type II report, whether they operated effectively over a period.

TCPA

The TCPA is the United States law restricting calls and texts made with autodialers or with an artificial or prerecorded voice, and since a February 2024 FCC declaratory ruling it plainly covers AI-generated voices.

Voice biometrics and anti-spoofing

Voice biometrics identifies or verifies a person by the characteristics of their voice, and anti-spoofing is the countermeasure that tries to work out whether the voice on the line is a real person or a synthetic copy.

Watermarking of synthetic audio

Watermarking of synthetic audio embeds an imperceptible signal into generated speech so it can later be identified as machine-made, and it is the main technical answer to the question of how anyone will know what was synthesized.

Voice navigation as accessibility

Voice navigation means moving through a website by speaking instead of clicking, which gives people with motor, vision or situational limitations another way in, but it is an additional modality rather than a route to accessibility compliance: it excludes anyone who cannot speak aloud, and it has to meet accessibility standards itself.