Realtime avatar terms, defined.
The words that come up when you compare avatar products, in plain language.
On-device avatar
A talking character whose face is animated and drawn by the user's own phone or computer, rather than rendered on a server and streamed as video. Nothing visual leaves the device, there is no per-frame server cost, and the number of people talking at once is not limited by anyone's GPU capacity. Yoob is an on-device avatar: the face runs on the iPhone Neural Engine at 25 frames a second, or on WebGPU in desktop Chrome and Edge.
Realtime avatar
A character that answers in a live conversation, with lips and expressions that follow its speech as it is produced, rather than a video generated ahead of time. Most realtime avatars render on a datacenter GPU and stream video to the viewer, billed per minute. A realtime avatar can also be drawn on the device, which is how Yoob works.
Lip-sync
Matching a character's mouth movements to speech. Good lip-sync is judged twice: while the character talks, and while it is silent. In silence the lips should rest closed; Yoob measures mouth closure in silence at 0.80, against 0.227 for the commercial lip-sync tool we trained against.
Viseme
The visual shape of the mouth for a group of speech sounds, such as the closed lips of p, b and m. Older animation systems map audio to a fixed set of visemes. Yoob's model instead predicts the face directly from audio features, 25 times a second, so it is not limited to a fixed set of shapes.
Neural Engine
The machine-learning accelerator in Apple's iPhone and iPad chips. It runs Core ML models at low power. Yoob's audio encoder and face renderer run on it, which is how a talking face fits inside about 32 MB and draws 25 frames a second without heating the phone the way GPU rendering would.
See also: Yoob vs cloud avatars · Device check · SDK docs