A new entrant in the consistent-character race
On October 21, 2026, Tokyo-based EmbodyMe announced the global expansion of DigiSelf, which it calls the fastest real-time AI avatar platform based on its own benchmarks against competing services. The pitch is different from the cinematic video models creators may already know: instead of generating polished scenes, DigiSelf is built for live, conversational avatars that respond to audio almost instantly.
For creators thinking about a recurring AI host or character, this matters because the underlying promise is the same one driving interest in tools like HeyGen and Kling: take a single photo, lock in a consistent face and voice, and reuse that character across unlimited new scripts without reshooting. DigiSelf is approaching that problem from the real-time, talking-avatar side rather than the cinematic-video side.
The speed numbers, and what they actually measure
EmbodyMe's benchmarks focus heavily on latency, the delay between input and an avatar's response. According to the company, DigiSelf began speaking incoming audio in a median of 0.33 seconds, compared with 0.62 seconds for the next-fastest competitor tested, about 47% faster. In live voice conversations, it began answering in a median of 1.17 seconds after a user stopped speaking, versus 1.94 seconds for the next-fastest competitor, roughly 40% less waiting time.
These figures come from EmbodyMe's own competitive testing, not an independent third party, so treat them as the company's claim rather than a verified industry standard. Still, for creators who've dealt with the awkward pause of an AI avatar visibly processing before it speaks, a sub-second response time is a real, noticeable difference if the claims hold up in practice.
The most striking number is for offline use: DigiSelf's local SDK reportedly begins speaking incoming audio in just 0.005 seconds when running entirely on a user's own hardware, with no round trip to the cloud at all.
Cloud API or offline SDK: two different trade-offs
DigiSelf offers two integration paths. The online API works like most cloud avatar tools, sending data to EmbodyMe's servers for processing. The local SDK runs entirely on a business's or creator's own hardware, with no user data sent to the cloud. Offline deployments are licensed as a one-time purchase with no per-minute usage fees, which is a meaningfully different cost model than the subscription or credit-based pricing most short-form tools use.
For creators who generate a high volume of short clips, or who are nervous about ongoing per-minute costs eating into a thin margin, a one-time license is worth watching even if it's aimed primarily at business customers right now. For anyone handling sensitive scripts, client information, or likeness rights they'd rather not route through a third-party server, the offline, no-data-leaves-the-device option is also worth noting as a category, separate from whether DigiSelf specifically fits a given budget.
How an avatar actually gets built
Creating an avatar requires a single photo plus about 10 to 30 seconds of voice, or alternatively a video of roughly 30 seconds to 2 minutes. DigiSelf then builds a dedicated AI model for that specific avatar, aiming to reproduce facial expressions, mouth movements, voice, speaking style, and mannerisms. Once built, new videos can be generated from new scripts without filming again.
This is the part that speaks most directly to the consistent-cast problem: the setup cost is small and happens once, and the output is a reusable character rather than a one-off clip. That's the same structural idea behind other avatar and consistent-character tools, just applied to a real-time, conversational product instead of a cinematic one.
DigiSelf isn't unproven. It launched first in Japan in March 2026 and has already been used by broadcasters, public figures, and government bodies, including Bunka Broadcasting, which worked with Dentsu to use an AI avatar of comedian Atsushi Tamura to read news on his radio show. EmbodyMe also says its apps have passed 2 million downloads across more than 200 countries and regions, with about 95% of that usage coming from outside Japan, which suggests the company has traction beyond a single market.
Where this fits for short-form creators
DigiSelf's pricing, as listed on its own site, includes a free Basic tier with 3-minute video generation and one custom voice model, a $59/month Personal tier with one custom avatar and 30 minutes of video per month for non-commercial use, and a Business tier with commercial use, API access, and real-time API access available by contacting sales. That structure suggests the free and Personal tiers are aimed at individuals experimenting or building personal content, while real commercial and real-time features sit behind a sales conversation.
The practical use cases EmbodyMe highlights, marketing, sales, training, customer service, reception, and digital signage, plus avatars usable in Zoom, Google Meet, and Microsoft Teams meetings, are mostly business-facing rather than built for short-form platforms like TikTok or Reels. That said, the underlying capability, a fast-responding, single-photo avatar that can deliver new scripts on demand, maps reasonably well onto talking-head intros, FAQ-style shorts, or dubbing a single recurring character into multiple languages without re-recording.
Whether DigiSelf's output quality, naturalness, and export options are actually suited to short-form editing workflows isn't something the brief confirms, so that's a real open question before anyone builds a workflow around it. What is clear is that the consistent-cast problem is being worked on from more than one direction, and real-time conversational avatars are now a second front worth tracking alongside cinematic video generators. If you're testing character-based workflows in Episodo, it's worth keeping an eye on tools like this as a potentially faster or cheaper option for specific formats, even if they're not a drop-in replacement for cinematic generation.