A plain-words glossary of voice AI
Voice AI has picked up a pile of jargon, most of it borrowed from three different trades: telephones, speech science and machine learning. You do not need any of it to run an agent. But if you are comparing tools or reading a contract, these are the words that come up, each in one plain paragraph. We have grouped them so you can skim.
The three jobs inside an agent
STT (speech-to-text)
The part that listens and writes down what the caller said. Also called a transcriber or ASR, for automatic speech recognition. Good ones handle accents, background noise and half-finished sentences. In NordTell you pick one from providers like Deepgram, Gladia or Soniox.
TTS (text-to-speech)
The part that reads the reply out loud in a chosen voice. This is where a call sounds human or robotic. NordTell uses voices from ElevenLabs and Cartesia, in every supported language.
LLM (large language model)
The part that thinks: it reads what the caller said and decides what to say back, when to use a tool, and when to hand over. The "brain" of the agent. You choose one from OpenAI, Anthropic or Google, and the provider is shown next to it.
How the three are put together
Pipeline
An engine that chains the three jobs in a row: listen, think, speak. The most control and the widest choice of voices, and usually the cheapest per minute. See our comparison of the engines.
Speech-to-speech (realtime)
One model that hears audio and speaks audio, with no written step in between. The most natural rhythm, because the model hears tone as well as words. Fewer voices, and it costs more.
Hybrid
A realtime model for the natural feel, with a second model checking every reply before it is spoken. It keeps the agent honest and on script. This is the engine we recommend. More in the product.
Words about the conversation
Barge-in
When the caller starts talking while the agent is still speaking, and the agent stops to listen, the way a polite person does. An agent that cannot be interrupted feels like a hold message.
Latency
The gap between the caller finishing and the agent replying. Measured in milliseconds. Too long and the call feels laggy; too eager and the agent talks over people. Realtime engines have the shortest latency.
Turn-taking
Deciding whose turn it is to speak. Humans use tiny cues, a breath, a falling tone. Voice AI has to guess when a caller has finished, not just paused to think.
Safety net
NordTell's own name for the phrases and checks that stop an agent from misbehaving in a given language: the words that mean "transfer me", the lines it falls back on, the rule that it cannot claim it booked something it did not. There are safety nets in thirteen languages. See how we stop an agent from bluffing.
Telephone words
SIP
Session Initiation Protocol, the standard that carries phone calls over the internet. If your provider "does SIP", you can bring your number to NordTell. See bringing your own number.
DID
A Direct Inward Dialing number, which is just industry-speak for a normal phone number that rings a specific line. When you buy or bring a number, you are dealing with a DID.
DTMF
The tones a phone makes when you press keys, as in "press 1 for sales". Voice agents let people speak instead, so DTMF menus are largely what they replace. See voice agents versus phone menus.
IVR
Interactive Voice Response, the old press-a-number phone menu. A voice agent is the conversational replacement: the caller just says what they want.
Carrier
The company that actually connects the call over the phone network, like Twilio or Telnyx. NordTell rides on top of a carrier; you can use ours or bring your own.
Warm transfer
Handing a call to a colleague the polite way: put the caller on hold, brief the colleague privately, then connect. The opposite is a cold transfer, where the call is thrown over the wall. See the difference and transfers in the product.
Words about connecting it to your tools
Tool
Something the agent can do during a call: check a calendar, look up an order, send a text. You describe it, point it at your system, and the model decides when to use it. More at tools and integrations.
Webhook
A message NordTell sends to your server the moment something happens, like a call finishing. Instead of you asking "anything new?", it taps you on the shoulder with the transcript and summary. For developers: the API and webhooks.
MCP
Model Context Protocol, a common way to expose a set of tools to an AI agent. If you run an MCP server, its tools appear in the agent editor without custom work.
API
The way to drive NordTell from your own code: create agents, place calls, upload leads, all over plain web requests. See developers.
Knowledge base
The documents and web pages the agent is allowed to answer from: your prices, opening hours, return rules. It quotes your text instead of guessing. More at knowledge base.
Words about running it responsibly
Consent
The caller's agreement to be recorded. In many places you must ask before recording. NordTell lets you set recording to off, on, or on-after-consent, per agent. See recording and consent in Europe.
Retention
How long you keep calls, recordings and transcripts before they are deleted. You set the number of days; NordTell deletes them on schedule.
Audit log
A record of who changed what and when in your workspace. Useful when more than one person has the keys, and often expected by a compliance review.
Disclosure
Telling the caller they are speaking to an AI. NordTell agents do this in their greeting, because it is honest and increasingly the law. See why we disclose.
Minutes
How NordTell is priced: you pay for minutes of conversation, not for seats or per agent. Each package includes a bucket of minutes and a price for extra ones. See pricing.
Words about the whole thing
Voice agent
The whole package: a named assistant with instructions, a voice, a language, a phone number and a set of tools, that answers or places calls on your behalf. Everything above is a part of one.
Inbound and outbound
Inbound is answering calls that come to you. Outbound is placing calls, one at a time or as a campaign. The same agent can do both.
Campaign
A list of people to call, a calling window and retry rules, run by the dialer. Used for reminders, renewals and follow-ups. See the outbound guide.
White label
Running NordTell under your own name, colours and domain, so your clients never see ours. For agencies and resellers. See white label.
Where NordTell fits
You can run a good agent knowing none of these words. They matter most when you are comparing tools or writing an integration, and then it helps that NordTell keeps them visible instead of hidden: the model, the voice, the retention and the consent setting are all things you can see and choose.