[

Vocabulary

]

Voice AI Buyer’s Vocabulary

What it means, what’s normal, what to ask. 60 Voice AI terms in plain English for founders, business leaders and buyers.

60 terms · 11 categories · Updated 1 Oct 2026

Supported by

InsurTech Club – The Insurance AI Community

CU Circle – The Credit Union AI Community

All

Basics

Hearing (Speech-to-Text)

Thinking (LLM)

Speaking (Text-to-Speech)

Conversation Flow

Speed & Quality

Phone & Telephony

Business Metrics

Compliance & Rules

Testing & Monitoring

Building & Buying

All terms

India-specific terms only

SORT

By category

A–Z

Basics

Voice AI agent

Software that talks to customers on a phone call like a human: it listens, understands, replies and takes actions.

Basics

IVR vs voice agent

IVR = Interactive Voice Response

IVR is the old "Press 1 for…" menu system. A voice agent understands normal speech, so callers can just say what they want.

Basics

Cascaded pipeline

The common way voice agents work, in three steps: speech is turned into text (STT), an AI model thinks (LLM), then text is turned into speech (TTS).

Basics

Speech-to-speech model

One AI model that listens and speaks directly, without converting to text in between.

Basics

Inbound vs outbound

Inbound means customers call you. Outbound means the AI calls customers.

Hearing (Speech-to-Text)

STT / ASR

Speech-to-Text / Automatic Speech Recognition

Speech-to-text, also called automatic speech recognition. It turns the caller's voice into text the AI can read.

Hearing (Speech-to-Text)

WER (word error rate)

Word Error Rate

The percentage of words the speech-to-text gets wrong. Lower is better.

Hearing (Speech-to-Text)

Code-switching

When a caller mixes languages in one sentence.

Hearing (Speech-to-Text)

Diarization

Working out who is speaking when, e.g. separating the customer from the agent, or noticing a second person in the room.

Hearing (Speech-to-Text)

Custom vocabulary / keyword boosting

Teaching the speech-to-text special words it should listen for, such as product names, brand names or local terms.

Thinking (LLM)

LLM

Large Language Model

The "brain" of the voice agent. It understands what the caller said and decides what to say next. Examples: GPT, Gemini, Claude, Llama.

Thinking (LLM)

System prompt

The written instructions that tell the AI who it is, what to do and what not to do.

Thinking (LLM)

Function calling

When the AI takes an action during a call, like checking a balance, booking a slot or updating the CRM.

Thinking (LLM)

RAG / knowledge base

RAG = Retrieval-Augmented Generation

Letting the AI look up answers from your documents, like FAQs, policy terms or product sheets, instead of guessing.

Thinking (LLM)

Hallucination

When the AI confidently says something that's wrong or made up.

Thinking (LLM)

Guardrails

Rules that stop the AI from saying or doing things it shouldn't.

Speaking (Text-to-Speech)

TTS

Text-to-Speech

Turns the AI's reply into a spoken voice.

Speaking (Text-to-Speech)

Voice cloning

Creating an AI voice that sounds like a specific real person.

Speaking (Text-to-Speech)

Prosody

The rhythm, stress and tone of speech, which makes a voice sound natural rather than robotic.

Speaking (Text-to-Speech)

Pronunciation dictionary

A list that tells the text-to-speech exactly how to say tricky words.

Speaking (Text-to-Speech)

Multilingual voice

One voice that can speak several languages, keeping the same personality.

Conversation Flow

Barge-in

When the caller interrupts the AI, and the AI stops talking to listen.

Conversation Flow

Turn detection

Deciding when the caller has finished speaking and it's the AI's turn.

Conversation Flow

VAD (voice activity detection)

Voice Activity Detection

Detecting when someone is speaking versus silence or noise.

Conversation Flow

Endpointing

How long the AI waits in silence before deciding the caller is done speaking.

Conversation Flow

Backchanneling

Small sounds like "hmm," "okay" or "ji" that show the AI is listening.

Conversation Flow

Silence handling / dead air

What the AI does when nobody speaks, or when it's waiting on a system.

Conversation Flow

Human handoff (warm transfer)

Passing the call to a human agent, along with a summary so the customer doesn't have to repeat themselves.

Conversation Flow

Voicemail detection (AMD)

AMD = Answering Machine Detection

Detecting whether an outbound call reached a person or a voicemail or recorded message.

Speed & Quality

Latency

The delay between the caller finishing speaking and the AI starting to reply.

Speed & Quality

Jitter and packet loss

Network problems that make audio choppy, robotic or missing.

Speed & Quality

Uptime / SLA

SLA = Service Level Agreement

How often the service is available, and what the vendor promises in writing.

Speed & Quality

8kHz audio

Standard phone call audio quality, which is lower than app or WhatsApp audio.

Phone & Telephony

SIP trunking

SIP = Session Initiation Protocol

A way to connect phone lines to your software over the internet, instead of physical lines.

Phone & Telephony

PSTN

Public Switched Telephone Network

The regular public phone network: normal mobile and landline calls.

Phone & Telephony

DID / virtual number

DID = Direct Inward Dialing

A phone number assigned to your AI agent that customers can call, or see when the AI calls them.

Phone & Telephony

WebRTC

Web Real-Time Communication

Technology for voice calls inside websites and apps, without going through the phone network.

Phone & Telephony

Concurrent calls

How many calls the AI can handle at the same time.

Phone & Telephony

Caller ID & spam labelling

The name or number customers see when you call, and whether their phone marks it as spam.

Phone & Telephony

DTMF

Dual-Tone Multi-Frequency

Keypad presses during a call ("Press 1…").

Business Metrics

Containment rate

The percentage of calls fully handled by the AI, without a human.

Business Metrics

AHT (average handle time)

Average Handle Time

The average length of a call.

Business Metrics

First call resolution (FCR)

FCR = First Call Resolution

Whether the customer's issue was solved in the first call, with no callback needed.

Business Metrics

CSAT

Customer Satisfaction Score

Customer satisfaction, usually a rating at the end of the call.

Business Metrics

Connect rate

The percentage of outbound calls answered by a real person.

Business Metrics

Right-party contact (RPC)

RPC = Right-Party Contact

Reaching the correct person, not a relative or a wrong number.

Business Metrics

Cost per minute

The total cost of one minute of AI calling: telephony, speech-to-text, LLM, text-to-speech and platform fees.

Compliance & Rules

DPDP Act

Digital Personal Data Protection Act

India's personal data protection law. It governs how personal data is collected, used and stored.

Compliance & Rules

TRAI calling rules & DND

TRAI = Telecom Regulatory Authority of India; DND = Do Not Disturb

Telecom rules on who can be called, when, and from which numbers. DND is the "Do Not Disturb" registry customers can opt into.

Compliance & Rules

140 and 160 number series

Special number series in India: 140 for promotional calls, and 160 for service and transactional calls from regulated businesses like banks and insurers.

Compliance & Rules

Consent & recording disclosure

Telling customers the call is recorded or handled by AI, and getting permission where needed.

Compliance & Rules

PII redaction

PII = Personally Identifiable Information

Hiding personal details, like card numbers, Aadhaar and phone numbers, in transcripts and recordings.

Compliance & Rules

Data residency

Where customer data and call recordings are stored and processed.

Testing & Monitoring

Evals

Evaluations

Structured tests that check how well the AI performs, before and after any change.

Testing & Monitoring

Simulated calls

AI-generated test callers that call your bot to test it at scale.

Testing & Monitoring

Call analytics / observability

Dashboards and tools that show what's happening on calls: errors, drop-offs, sentiment and latency.

Building & Buying

Build vs buy

Deciding whether to build your own voice agent or buy a ready-made product.

Building & Buying

POC / pilot

POC = Proof of Concept

A small trial run before a full contract.

Building & Buying

Usage-based pricing

Paying by usage, such as per minute or per call, instead of a fixed fee.

Building & Buying

Vendor lock-in

When it's hard or expensive to switch to another vendor later.

An independent guide by Unio. Benchmarks are typical ranges and are updated as we collect data from the community.

Voice AI Buyer’s Vocabulary

Basics

Voice AI agent

IVR vs voice agent

Cascaded pipeline

Speech-to-speech model

Inbound vs outbound

Hearing (Speech-to-Text)

STT / ASR

WER (word error rate)

Code-switching

Diarization

Custom vocabulary / keyword boosting

Thinking (LLM)

LLM

System prompt

Function calling

RAG / knowledge base

Hallucination

Guardrails

Speaking (Text-to-Speech)

TTS

Voice cloning

Prosody

Pronunciation dictionary

Multilingual voice

Conversation Flow

Barge-in

Turn detection

VAD (voice activity detection)

Endpointing

Backchanneling

Silence handling / dead air

Human handoff (warm transfer)

Voicemail detection (AMD)

Speed & Quality

Latency

Jitter and packet loss

Uptime / SLA

8kHz audio

Phone & Telephony

SIP trunking

PSTN

DID / virtual number

WebRTC

Concurrent calls

Caller ID & spam labelling

DTMF

Business Metrics

Containment rate

AHT (average handle time)

First call resolution (FCR)

CSAT

Connect rate

Right-party contact (RPC)

Cost per minute

Compliance & Rules

DPDP Act

TRAI calling rules & DND

140 and 160 number series

Consent & recording disclosure

PII redaction

Data residency

Testing & Monitoring

Evals

Simulated calls

Call analytics / observability

Building & Buying

Build vs buy

POC / pilot

Usage-based pricing

Vendor lock-in

Get new Voice AI terms and benchmarks every month

Heard a Voice AI word you didn’t understand?