Editing Kismet (robot) (section)

===Low-level feature extraction system===
This system processes raw visual and auditory information from cameras and microphones. Kismet's '''vision system''' can perform eye detection, motion detection and, albeit controversial, skin-color detection. Whenever Kismet moves its head, it momentarily disables its motion detection system to avoid detecting self-motion. It also uses its stereo cameras to estimate the distance of an object in its visual field, for example to detect threats—large, close objects with a lot of movement.<ref name="Kismet, the robot">{{cite web|title=Kismet, the robot|url=http://www.ai.mit.edu/projects/sociable/baby-bits.html}}</ref>

Kismet's '''audio system''' is mainly tuned towards identifying affect in [[infant-directed speech]]. In particular, it can detect five different types of affective speech: approval, prohibition, attention, comfort, and neutral. The affective intent classifier was created as follows. Low-level features such as pitch mean and energy (volume) variance were extracted from samples of recorded speech. The classes of affective intent were then modeled as a [[gaussian mixture model]] and trained with these samples using the [[expectation-maximization algorithm]]. Classification is done with multiple stages, first classifying an utterance into one of two general groups (e.g. soothing/neutral vs. prohibition/attention/approval) and then doing more detailed classification. This architecture significantly improved performance for hard-to-distinguish classes, like ''approval'' ("You're a clever robot") versus ''attention'' ("Hey Kismet, over here").<ref name="Kismet, the robot"/>