Glossary of common terms

Category (face-to-face, agentic)

Human computing: The movement to close the gap between people and machines by making technology feel unmistakably human. Human computing replaces traditional interfaces with natural human interaction, letting computers see, hear, understand, remember, respond, and even look like we do. Tavus is “a research lab pioneering human computing.”

PAL (Personified Application Layer): A new kind of application you talk to instead of operate. A PAL learns from us, collaborates with us, and develops an understanding of us, until it becomes a trusted assistant, sidekick, companion, or coworker. Under the hood, a PAL holds the full configuration for its role: behavior, knowledge, objectives, guardrails, tools, and every pipeline layer, while its Face supplies the on-screen likeness and voice. PALs are the through line of every Tavus product, whether built in PAL Maker, driven through the CVI API, or deployed by Solutions, and every one runs on the same foundational models: Phoenix-4, Raven-1, and Sparrow-1. Use “PAL” as the default external term for what Tavus builds. In the API, PALs appear as pal_id, and the legacy persona_id remains supported.

The translation tax: Using a machine has always meant learning its language: the commands, then the menus, now the prompt. The cost was never the idea, it was turning the idea into the machine’s terms before it could help you. Human computing removes that tax, so machines learn our language instead of us learning theirs.

Teaching machines the art of being human: The Tavus mission: teach machines the art of being human, so that interacting with one feels as natural as talking to a friend or coworker. Appears in short blurbs and company descriptors.

Humanlike presence: What Tavus gives your product: lifelike presence that deepens engagement, trust, and feedback. PALs bring emotional intelligence to every conversation, at scale.

Products

Conversational Video Interface (CVI): The developer framework for real-time, face-to-face conversation, and the world’s fastest interface of its kind. CVI maps a Face (the visual) and a PAL (the behavior) onto a live conversation, running the full pipeline of perception (Raven), conversational flow (Sparrow), speech recognition, language, speech synthesis, and real-time rendering (Phoenix) over WebRTC with sub-600ms utterance-to-utterance latency. Plug in your own components where you want control, and everything PAL Maker does is available programmatically through the same API.

PAL Maker: Where anyone builds a PAL, with no code. Describe the PAL you want in plain words and Charlie, the kid in the machine who lives inside PAL Maker and builds alongside you, configures the whole thing: the system prompt, Face, objectives, guardrails, tools, knowledge, and advanced settings, ready to talk and ready to go to work before you leave the page. Fork a starter template into your account or open the GitHub repo behind it, then publish straight to a widget, an embed, or a hosted landing page. Always two words, “PAL Maker.” Lives at maker.tavus.io.

Solutions: Enterprise deployments of PALs, designed, built, tuned, and managed by the Tavus team alongside yours. Production-ready and scalable across go-to-market, healthcare, learning and development, and recruiting: a PAL rep, a PAL coach, or a PAL intake assistant, live in weeks, with custom PAL design, performance tuning, and support for complex workflows, and no infrastructure for your team to stand up.

Face (Personal / Stock): The rendered on-screen identity a PAL uses, supplying its likeness and voice, brought to life by Phoenix-4. “Personal” is trained from a creator’s footage, from a single image or two minutes of video. “Stock” is a pre-built library of optimized identities ready for immediate use. Fully white-labeled, fast training, identity preservation. “Face” replaces “replica” in Tavus copy. In the API, Faces appear as face_id, and the legacy replica_id remains supported.

Core models (the “Tavus brain”)

Phoenix-4 (face rendering): Full-face, real-time generation with every pixel rendered live, pixel-level identity preservation, studio-grade lip sync, and micro-expressions for true realism. Built-in emotional intelligence lets it show genuine emotion, and full-duplex operation means it listens and responds at the same time. Powers Faces in CVI and video generation.

Raven-1 (multimodal perception): Real-time audio and visual understanding. Raven-1 fuses tone, prosody, facial expression, posture, and gaze into a single representation a language model can reason over, maintains ambient awareness, detects specified gestures and objects, and can trigger perception tools based on what it sees and hears. Works with screen share and other visual inputs.

Sparrow-1 (conversational timing): Transformer-based conversation timing that adapts to human rhythm across tone, pauses, and overlap. Sparrow-1 models the conversational floor continuously, resolves interruptions within tens of milliseconds, and adapts to individual speaking patterns within a single session with no calibration needed.

Capabilities and platform primitives

Perception: Real-time reading of facial cues, environment, and on-screen content to inform responses, capture emotional and behavioral signals, and trigger actions. (Model: Raven-1.)

Turn-taking: Natural conversational flow: when to listen, when to interject, when to continue, so sessions feel human. (Model: Sparrow-1.)

Skills: Pre-built capabilities authored by Tavus that you attach to a PAL from the skill registry: Internet Search, Presentation Mode, and Magic Canvas today, with more added over time. Each skill bundles the prompting, configuration, and wiring a capability needs, and it is active the moment it is attached.

Magic Canvas: A skill that lets a PAL show interactive cards during a live video conversation: multiple-choice questions, typed inputs, date and time pickers, live scheduling embeds, formatted text, charts, and alerts, seven components in all. The PAL decides when a card appears, cards render in a side rail beside its video, and every response flows back to the PAL and to your webhook. Steer when cards appear with usage guidance, the system prompt, or per-conversation context, and render with Tavus-designed cards or your own React components. Not available in audio-only, text-chat, or external-meeting conversations such as Google Meet.

Presentation Mode: A skill that lets a PAL present PDF decks and images straight from your Knowledge Base, in two ways: Walk the Deck, where the walk-through drives the conversation, and On-Demand, where the PAL pulls up the relevant slide when the conversation calls for it. Slides stream as the PAL’s own screen share, presentation style is steerable with a custom prompt, and PDFs up to fifty pages are presentable.

Meetings (Google Meet): The conferencing layer gives a PAL its own invitable email identity, so it can be added to a Google Meet over Google Calendar like any teammate. An allowlist controls who may invite it, it joins scheduled meetings about a minute before the start or joins ones already in progress, and it participates with its Face’s likeness and voice. Zoom and Microsoft Teams are coming soon.

Tool Calling: Lets a PAL trigger code while the conversation is happening: look something up, write to a CRM, hit a third-party API, or notify your frontend. Tools are reusable objects, created once and attached to any number of PALs, in two in-call types: LLM tools triggered by what the user says, and perception tools triggered by what Raven sees or hears. Post-call actions run once after the conversation ends, filled in from the transcript and perception analysis. Results are delivered as an app message to your frontend or as a direct API call.

Internet Search: A skill that lets a PAL answer with up-to-date information from the web. With it attached, the PAL searches in real time on every turn and grounds its response in the latest information. A pure on and off toggle with no configuration.

Knowledge Base (RAG): Document grounding for accuracy and recency. Upload files (.pdf, .txt, .docx, .doc, .png, .jpg, .pptx, .csv, and .xlsx) or public website URLs, group documents with shared tags, and choose a speed, balanced, or quality retrieval strategy. The PAL continuously analyzes the conversation and pulls the relevant information in as added context. Manage it in PAL Maker or through the API.

Memories: Persistent memory across sessions so conversations pick up where they left off. Memories form during a conversation and are referenced in later ones, organized with memory stores, a flexible tag system that keeps each participant’s history with each PAL in its own bucket.

Objectives: Goal-oriented instructions that define the desired outcomes and flow of a conversation, working alongside the system prompt. Plan an entire workflow as a branching structure, gathering contact information, understanding budget, assessing decision-making authority, and the PAL carries participants from start to finish while keeping the conversation natural.

Guardrails: Strict behavioral guidelines a PAL follows throughout every conversation, working as a safety layer alongside the system prompt. Guardrails both steer the PAL and flag violations: each one is evaluated continuously in the background while the conversation runs, and a violation triggers a real-time callback or app message so you can react. Each guardrail is its own resource, created, attached, and edited independently.

Starter templates (Stock PALs): Complete, live PALs you can fork into your account with no code, an interviewer, an SDR, a medical intake agent, a casting director, matching the presets in PAL Maker, with the GitHub repo behind each one available when you want the code.

Deployments (Widget, Embed, Landing page): Hosted, pre-configured experiences published from PAL Maker and dropped into any site with no backend and no API key in the browser. Three channels: a Widget, a floating launcher added with one script tag and one element, an Embed, an inline surface that feels native to the page, and a Landing page, a standalone Tavus-hosted page you can share as a link. Usage limits and abuse protection are handled server-side, and Magic Canvas cards render automatically in hosted widgets and embeds.

Bring your own LLM (BYO LLM): Swap in your preferred language model to control tone, cost, or domain expertise while keeping Tavus’s humanlike presence.

White-labeling: APIs and endpoints designed so the experience is fully your brand, including white-labeled Face consent flows at higher tiers.

Alpha-channel video: Output with transparency for compositing onto any background or UI.

Transcripts and recordings: Conversation logs and optional recordings for analytics, QA, or compliance, with per-minute pricing where applicable.

Concurrency / streams: How many simultaneous live conversations your plan supports (for example, 1 → 3 → 15 → custom).

Canonical use-case nouns (how customers talk about outcomes)

Digital twin (customer phrasing): A lifelike, interactive stand-in for a person (expert, coach, creator) used in real-time conversations. Common in customer language. In Tavus copy, describe it as a PAL built on a personal Face.

AI interviewer / mock interviews: Face-to-face interview practice and live first-round screening with feedback and scorecards. A flagship use case for CVI and for recruiting teams. In Tavus copy: PAL interviewer or PAL recruiter.

AI SDR: Engages website visitors and prospects in live, face-to-face sales conversations, qualifying, answering questions, and booking meetings. A flagship go-to-market use case. In Tavus copy: PAL rep.

AI intake assistant / medical intake: Patient intake and navigation over a face-to-face call, reading how someone is feeling and meeting them in their language. A flagship healthcare use case. In Tavus copy: PAL intake assistant.

AI tutor / coach: Role-play, upskilling, and personalized instruction that adapts in real time to learner signals, from B2B coaching with tight guardrails to consumer study-buddy Knowledge Base use cases. A flagship learning and development use case. In Tavus copy: PAL coach or PAL tutor.

AI companion: Humanlike support for wellbeing and daily check-ins (for example, senior care and loneliness mitigation) with lifelike presence and low latency. In Tavus copy: PAL companion.

‍