AI call center and speech analytics glossary

Terms by topic

AI call center

A system in which artificial intelligence answers and places calls, replies in chats and scores agents' conversations. People stay for complex cases: the AI hands the conversation over. In Leksta, calls, chats and analytics share one console.

Voice bot (AI voice agent)

Software that holds a phone conversation by voice: it recognizes the caller's speech, composes an answer and speaks it with a synthetic voice. Unlike IVR, it understands free speech, not only key presses. The Leksta voice bot speaks Russian and says it is an AI.

IVR (interactive voice response)

A PBX voice menu: "press 1 to…". It follows a fixed tree and mostly understands key presses and pre-recorded phrases. A voice bot differs because it understands free answers and replies from a knowledge base.

Speech analytics

Automatic analysis of call recordings: transcription, keyword search, scoring against criteria, sentiment and call goals. It lets you review every call instead of a sample. In Leksta, the AI scores every analyzed call.

Transcription

Turning a call recording into time-stamped text. It is the basis of speech analytics: phrases are searched and the call is scored on the text. On Russian telephony, names and numbers may be misheard, so key fragments are checked against the audio.

Diarization (speaker separation)

Deciding who said each line: the agent or the customer. It works best on dual-channel recordings where voices go to separate channels. On mono recordings some lines may stay with an unknown speaker.

RTF (real-time factor)

Recognition speed ratio: processing time divided by audio length. RTF 0.5 means one minute of audio is processed in 30 seconds. In an early test-stand measurement, Leksta transcription showed RTF 0.27–0.37; this is a measurement, not an SLA.

DTMF (touch tones)

The signals a phone sends when keys are pressed. A voice bot uses them for unambiguous answers: "press 1 to confirm the appointment". A key press is recognized more reliably than a short "yes" in a noisy place.

AMD (answering machine detection)

Detecting whether a person or a voicemail picked up. In outbound campaigns it stops the bot from talking to voicemail: the bot leaves a short message or ends the attempt, and the retry follows the schedule.

SIP

Session Initiation Protocol, the standard IP telephony protocol for setting up and ending calls. PBXs, carriers and voice bots connect over SIP without special hardware.

SIP trunk

A SIP connection to a carrier that carries a company's inbound and outbound calls. It lets a voice bot work on your existing number: the number and the carrier stay yours.

Virtual PBX (cloud PBX)

A carrier-hosted phone system: numbers, call routing and call recording without office hardware. A1 VATS in Belarus is an example. Leksta receives calls and recordings from it for analytics.

IP PBX (Asterisk)

A software phone system installed on a company server. Asterisk is the most widely used open-source IP PBX. Leksta connects to an existing Asterisk without changing the number.

FreeSWITCH

An open-source telephony platform for call routing, conferencing and recording. Leksta uses it for outbound campaigns and for the voice bot to take calls.

Omnichannel

An approach where every customer channel — calls, website chat, messengers — flows into one system with a shared history and shared answers. Customers don't repeat themselves when they switch channels.

Shared inbox

One window that collects conversations from every connected chat channel. Agents see the thread, assign owners, set statuses and reply with templates. In Leksta it is the Chats section next to Calls.

Knowledge base

The company documents — policies, prices, services, FAQs — that the AI answers from. Update a document and the answer changes in both voice and chat.

RAG (retrieval-augmented generation)

A method in which the AI first finds relevant passages in the knowledge base and then writes the answer from them. It reduces made-up answers and makes it possible to show the source.

LLM (large language model)

A neural network trained on large amounts of text that understands a question and answers in natural language. In a contact center it drafts answers, call summaries and checklist scores.

Checklist scoring

Checking a conversation against set criteria: greeting, needs discovery, objection handling, outcome. The AI marks each item and explains the mark with a quote from the call. The company defines the criteria.

Score calibration

Comparing AI scores with a manager's scores on the same calls. It helps refine criteria wording so the AI and people read each checklist item the same way.

Sentiment

The emotional tone of speech: calm, irritated, satisfied. In Leksta it is detected for the customer only. Agent mood and emotions are not assessed.

Keyword dictionary

A list of words and phrases to find in calls: "complaint", "cancel", a competitor's name, banned expressions. A match flags the call and can trigger an alert to a manager.

Handoff to an agent

Passing a call or chat from the AI to a person: when the customer asks, when the question is complex, or by your rules. The agent sees the history and doesn't re-ask.

Whisper coaching

A mode in which a supervisor or the system prompts the agent during a call without the customer hearing it. Used to train new staff.

Outbound campaign

Calling a contact list on a schedule: reminders, confirmations, surveys. You set the calling window, time zone, number of attempts and pauses between them. Each call's result is recorded.

Call retries

Rules for repeat attempts when the person didn't answer, was busy or voicemail picked up: how many times, with what pause and at what hours.

Do-not-call list (opt-out)

Numbers that must not be called because the person opted out or gave no consent. In Leksta, pressing 9 during a call blocks the number immediately, and the do-not-call list always overrides an uploaded list.

AI disclosure

Telling the customer at the start of a call or chat that they are talking to artificial intelligence. It is honest practice and a requirement of some regulators. The Leksta voice bot introduces itself as an AI.

On-premise

Installing software on the customer's own servers instead of the vendor's cloud. Data, recordings and transcripts stay inside the company. Leksta runs this way on test machines, including Astra Linux SE 1.8.

Air-gap (closed network)

Operating with no connection to outside networks: the system does not reach the internet or cloud services. In Leksta it was checked by network capture on a test stand — no packet left the allowed set of addresses.

Local AI model

A language model and speech recognition running on the customer's servers without calling the cloud. Hardware — CPUs, GPUs and memory — has to be sized for the load; sizing is done during the pilot.

PII masking

Automatically replacing phone numbers, card numbers and other personal data in transcripts and summaries with masks. It reduces leak risk when viewing and exporting. It applies to new data, not retroactively.

Webhook

An automatic HTTP notification a system sends to an outside service when an event happens: a call ended, a chat arrived. Through webhooks and n8n, Leksta events reach your systems.

CRM integration

Linking the contact center to the customer database: the call outcome, summary and recording appear on the customer record. Leksta writes results to Bitrix24 and amoCRM after the call; 1C is a pilot on a test database.

Per-minute pricing

A pricing model where you pay for minutes actually processed, not for agent seats. In Leksta, minutes above your plan's bundle are billed separately (0.06 to 0.09 BYN); see the pricing page for plans.

Read next

• Voice bot →How the voice bot handles inbound and outbound calls.
• Speech analytics →Transcription and checklist scoring of 100% of calls.
• Integrations →A1 VATS, Asterisk, FreeSWITCH, SIP, Bitrix24, amoCRM.
• On-premise →Installation on your servers and air-gapped mode.
• What is an AI call center →An in-depth article on how it works.