Voice AI glossary

The words used across voice AI, telephony, and Adam's own dashboard - what they mean, and what Adam does with them.

Voice and speech

The audio machinery every AI phone call runs on.

ASR (automatic speech recognition): The system that turns a caller's speech into text as they speak. Also called speech-to-text. Its accuracy caps everything downstream, measured by word error rate on real call audio. It works in real time, producing words before the caller finishes the sentence - and it's the first thing we tune for each language, Polish hardest.
STT (speech-to-text): The same job as ASR, under the name cloud platforms tend to use. On Adam it sits inside the all-in per-minute price - no separate speech vendor to wire up or invoice.
TTS (text-to-speech): The system that turns Adam's words into audio on the line. Judged on naturalness, prosody, and how fast the first sound arrives. Each language gets a native voice tuned for it, not a generic one bolted on.
Voice cloning: Building a synthetic voice from a short recording of a real speaker. Adam doesn't clone your staff - each language gets its own native voice, so Polish sounds Polish and not translated.
Latency: The gap between the caller finishing a sentence and Adam starting his reply. Under roughly 400 milliseconds it feels like a normal conversation; past a second, callers notice the delay and start to disengage. Holding it down drives most of the engineering in a voice agent.
Barge-in: When the caller talks over Adam mid-sentence. A production agent has to catch it within about 150 milliseconds, stop cleanly, and treat the interruption as the real input. People interrupt phone calls constantly, so this is table stakes.
VAD (voice activity detection): The component that decides when the caller has finished speaking. Tuning it trades responsiveness - cutting in early - against patience - sitting through a mid-sentence pause. Wrong tuning is the most common reason agents talk over callers.
Turn detection: The wider judgment of whose turn it is to speak - VAD plus the context of what's been said. Good turn-taking is the difference between a conversation and two people talking past each other.
Diarization: Separating who said what when more than one person is on the line. It keeps the transcript straight and tells Adam which voice to answer.
WER (word error rate): The standard measure of speech-recognition accuracy: the words misheard, dropped, or invented, over everything said. Lower is better. Noisy lines and proper names push it up; clean audio keeps it in the single digits.

Telephony

How a call reaches a phone.

PSTN (public switched telephone network): The global phone network every mobile and landline connects to. Adam's calls are ordinary calls on it: a real phone rings, no app needed.
SIP (session initiation protocol): The protocol phone systems use to set up, manage, and end calls over IP. The signalling layer under modern telephony.
SIP trunking: Routing calls through a SIP carrier connection instead of physical phone lines. How a business plugs its phone system into the wider network, and how a platform routes through an existing carrier.
E.164: The international phone-number format: a plus sign, country code, and number, no spaces - +48512345678. The format Adam's lead imports and API expect.
Caller ID: The number that shows on the phone as it rings. Adam calls from real, dedicated numbers, not anonymized or spoofed ones.
DTMF: The touch-tone signals a phone keypad sends - the 'press 1 for sales' codes that legacy menus run on.
IVR (interactive voice response): The menu tree older phone systems answer with - 'press 1 for sales'. Conversational agents like Adam replace it with a real conversation that resolves what the caller wants.
Answering machine detection: Telling a live 'hello?' from voicemail in the first seconds of a call. It decides whether Adam has a conversation or schedules a polite retry.
Warm handoff: Passing a caller to a human with the context attached, not a cold transfer. Adam's version is the hot queue: the lead, the transcript, and why they're hot, waiting for a human callback.
Concurrent calls: How many calls run at the same time. It sets how fast a campaign clears a list - leads worked in parallel instead of one after another.

The Adam platform

The words on Adam's own dashboard.

Pathway: The designed conversation: what Adam opens with, what he asks, how he reacts, when he ends the call. A new use case is a new pathway, not a new product.
Outcome: How a call ended, in one field: qualified, callback booked, not interested, no answer. Outcomes are what you filter, count, and route on.
Extraction: Turning the conversation into fields you can act on: still selling - yes; timeframe - next month; price - 30,000. Every call returns its answers as data, next to the transcript and the recording.
Guard rail: A rule that fixes what Adam may and may not do on a call - introduce himself as an AI, stay on topic, stop the moment someone opts out. Enforced by the platform, not left to the script.

Compliance and privacy

The rules AI calling lives under in the EU - and how the platform holds them.

GDPR: The EU's data-protection regulation. For call data, our customer is the controller - it's their leads - and Adam Calling is the processor; the platform side of that split (retention, deletion, export) is built in.
Controller and processor: Who decides why data is processed (the controller - our customer) versus who processes it on their instructions (the processor - us). The distinction GDPR hangs every duty on.
EU AI Act: The EU's AI law. Among other things, it requires that people know when they're talking to an AI - so Adam introduces himself as one at the start of every call.
Consent and opt-in: Adam calls people who agreed to be contacted. Consent lists and do-not-call registries are checked before dialing, not after complaints.
Do-not-call registry: A register of numbers that must not receive marketing calls. Checked before a campaign dials.
Data residency: Where data physically lives. For Adam: compute, storage, and database in AWS Frankfurt; telephony and voice processing homed in Europe; post-call analysis on EU model endpoints.
Retention: How long call data is kept. Recordings default to 90 days, the schedule is yours, and export or purge is always available.
Encryption: Recordings and data are encrypted at rest in Frankfurt and in transit everywhere. Standard, verifiable, boring - the way security should be.
DPA (data processing agreement): The contract that pins down the processor duties above. Available on request: hello@adamcalling.com.