Tavus

Tavus is a San Francisco-based real-time conversational digital human platform that lets an AI talk with you over video using its own face and voice, instead of playing back a pre-rendered clip. The foundation is the CVI real-time conversation pipeline: the digital human is called a PAL and its appearance a Face, while the retired Persona and Replica names are no longer used.

Core capabilities

• Sub-second conversation: three in-house models — Phoenix for rendering, Raven for perception, Sparrow for turn-taking — chained into one real-time pipeline.
• Face cloning: the PAL carries behavior and knowledge, the Face carries appearance; a single photo or a 1–2 minute video is enough.
• Real meeting attendance: a PAL has its own email identity, auto-accepts calendar invites, and joins Zoom, Meet or Teams 1 minute early.
• Developer-first: REST, OpenAPI, React components and zero-backend embedding, with the option to swap in LiveKit or your own LLM.
• No-code entry point: PAL Maker is a conversational builder, so non-technical colleagues can assemble a receptionist too.
• Compliance switches: the eu policy turns on identity disclosure and narrows emotion recognition in one step, backed by SOC 2 and HIPAA.

Where it fits Teams making an AI interviewer, sales role-play, medical follow-up or in-app face-to-face support the core of their product; scenarios where a digital human must actually join a Zoom call with customers or candidates; and English-first teams in North America and Europe. It is a poor fit for batch marketing video production, for Chinese knowledge-base Q&A, or for anyone who just needs a talking head for scripted voice-over.

Unique advantages Among peers, only Tavus treats real-time conversation as the foundation rather than a bolt-on feature: HeyGen's LiveAvatar is a separate sub-platform, Synthesia is strongest at enterprise pre-rendered courses, and D-ID is cheaper but stays image-driven. Comparing unit prices is pointless — real-time streams and pre-rendered video are not the same category.

Editorial take The technology deserves a POC; commercially, keep some skepticism. Review samples on all three platforms are tiny and contradictory: G2's three reviews average 3.0, Trustpilot's five reviews score zero, and Product Hunt's six reviews are all perfect — not enough to conclude anything; the official latency story has two versions, 500ms and 600ms, and the docs give no number; 42 languages include Chinese, yet the knowledge base only supports English, which blocks Chinese Q&A; the a16z backing cannot be substantiated — it is actually a Series B of 40 million dollars led by CRV, plus YC. My advice: run one real conversation on the free tier's 20 minutes first.