Digital Human

A digital human is a persistent, interactive, AI-driven avatar — combining a visual presence (photorealistic or deliberately stylized), real-time voice synthesis, lip sync, and a conversational AI reasoning backend (typically an LLM with retrieval-augmented context about a specific domain or persona) — designed to hold live, two-way conversations with a user rather than simply deliver a pre-scripted, one-way video. This distinguishes a digital human from avatar-generation-for-video products: avatar generation typically produces a finished, non-interactive video asset from a fixed script, while a digital human runs live, generating its responses, voice, and facial/lip animation in real time as a conversation unfolds — architecturally it's a full pipeline of speech-to-text (hearing the user) → LLM reasoning (deciding what to say) → text-to-speech (generating the voice) → lip sync/facial animation (rendering the response), all orchestrated with tight latency budgets since a conversational partner that pauses for 5+ seconds between turns feels broken. Platforms like Soul Machines, NVIDIA's ACE, and D-ID's Agents package this real-time pipeline as an embeddable widget or SDK, handling the difficult latency-orchestration work internally so a product team can drop in a working conversational avatar without independently building and tuning the STT/LLM/TTS/animation pipeline and its associated infrastructure themselves. Why it matters for SaaS builders: digital humans are the emerging interface layer for customer service (a visual, conversational alternative to a text chatbot), virtual sales/brand ambassadors on e-commerce sites, and interactive training simulations (practicing a difficult conversation with an AI-driven "customer" or "patient" persona). The engineering challenge is materially harder than any single output type covered elsewhere in this glossary because of the real-time, low-latency, multi-model-orchestration requirement — most builders integrate a vendor's full digital-human SDK rather than assembling the STT/LLM/TTS/lip-sync pipeline themselves for a live product. A concrete worked example — a retail SaaS adding a "virtual stylist": (1) a shopper opens the site's chat widget and it loads a digital-human avatar instead of a text-only bot; (2) the shopper speaks (or types) "I need an outfit for a summer wedding"; (3) STT (if voice) converts the query to text, which is sent with the store's product catalog as retrieval context to an LLM; (4) the LLM's text response streams to a low-latency TTS engine while simultaneously driving the avatar's lip-sync and expression in real time, so the avatar visibly "speaks" its product recommendations back to the shopper within roughly 1-2 seconds of the query.