EmbodyMe takes DigiSelf global, claiming the fastest real-time avatar response times

EmbodyMe, a Tokyo generative AI company founded in 2016, announced the global expansion of DigiSelf on 21 October 2026. The platform builds a speaking avatar from a single photo plus roughly 10-30 seconds of voice, or from a 30-second to two-minute video, and supports multilingual video, audio and live conversation inside Zoom, Google Meet and Microsoft Teams.

EmbodyMe reports median latency of 0.33 seconds to begin speaking incoming audio, against 0.62 seconds for what it calls the next-fastest competitor, and 1.17 seconds median response in live voice conversation against 1.94 seconds. Its offline SDK, which runs on the customer's own hardware with no user data sent to the cloud while running, is reported at 0.005 seconds to begin speaking incoming audio. Paid plans with real-time conversation start at $29 per month ($24 billed annually); API real-time avatar streaming starts at $0.10 per minute; the offline SDK is a one-time licence from $9,900. The company cites Japanese deployments since March 2026 including Bunka Broadcasting using an avatar of comedian Atsushi Tamura to read news on his radio show with Dentsu, an avatar of announcer Keitaro Sunayama filling in while he covered the Olympics, and Shizuoka Prefecture using an avatar of Governor Yasutomo Suzuki to open TECH BEAT Shizuoka.

Why it matters for digital humans: The on-premise angle is the part worth watching. Every benchmark here is company-run, with competitors and methodology unnamed, so the "world's fastest" claim should be treated as promotional until someone reproduces it. But a one-time licence for a locally hosted avatar runtime is a different commercial shape from per-minute cloud streaming, and it answers the data-residency objection that stalls avatar deployments in regulated sectors. If local inference can hold sub-second turn-taking on commodity hardware, the latency argument for cloud rendering weakens.