For the uninitiated, GLaDOS, the Genetic Lifeform and Disk Working System, is an iconic character from the critically acclaimed online game collection Portal, developed by Valve. Initially launched because the central antagonist within the unique Portal sport launched in 2007, GLaDOS rapidly grew to become certainly one of gaming’s most memorable and chilling villains. Created by Aperture Science, she serves because the controlling intelligence behind the Aperture Science Enrichment Middle, the place gamers navigate a collection of check chambers armed solely with a portal gun.

What makes GLaDOS notably intriguing is her complicated persona, which blends a chilly, calculating demeanor with darkish humor and occasional hints of vulnerability. Voiced with sinister attraction, GLaDOS’s dialogue is laced with sarcasm and veiled threats, making her interactions with the participant each unnerving and charming. Her distinct robotic voice and memorable traces, similar to “The cake is a lie,” have change into synonymous with the Portal franchise.

A villainous robotic like GLaDOS may appear to be an odd selection for a companion, however I suppose isolation could make an individual do some unusual issues. That was the case for {hardware} hacker and YouTuber Dave Niewinski, anyway, who got here down with COVID not too long ago and needed to preserve to himself for some time. Throughout that point, he determined to construct a sensible clone of GLaDOS to maintain him firm. And we’re glad he did — the interactive animatronic robotic is powered by AI to make it a great companion, even when it might be secretly bent on his destruction.

There are lots of items to a mission like this, however getting the voice proper is probably a very powerful think about making a convincing reproduction of GLaDOS. Niewinski did a fast survey of pre-built voice mills that search to copy GLaDOS’s voice, however discovered all of them to be missing, so a customized resolution was wanted. He discovered that Valve has numerous audio clips of GLaDOS talking on their web site, which made it simple to gather a dataset of over an hour in size.

This dataset was leveraged to coach a pair of machine studying fashions. The primary is a spectrogram generator known as Fastpitch, which was fine-tuned to sound extra like GLaDOS’s voice. The outcomes had been shut, however not fairly proper, so one other mannequin, known as a vocoder, was additionally educated to sound just like the goal voice and additional course of the outcomes. This mix of fashions produced a reasonably convincing GLaDOS voice.

These networks had been then loaded into NVIDIA’s Riva AI software program improvement package to make them simple to entry over a community. As an additional advantage, Riva additionally equipped speech-to-text capabilities that allowed folks to additionally speak to the robotic. The textual content generated by this pipeline is fed right into a locally-running giant language mannequin, and the outcomes it produces are then fed into the GLaDOS voice generator for a spoken response.

The fashions all ran regionally on a Forge Service Board for the highly effective but small NVIDIA Jetson AGX Orin pc module. This little beast of a pc even had sufficient processing energy left over to regulate a Unitree Z1 robotic arm that served as the bottom of the robotic. The rest of GLaDOS’s physique consisted of 3D-printed components and a few RGB LEDs. As a of entirety, to present the robotic extra persona, a Stereolabs ZED 2i depth digital camera was included within the construct. This fed information into an individual monitoring algorithm (sure, additionally working on the Jetson AGX Orin), and moved the Z1 robotic arm to maintain GLaDOS “trying” on the particular person it was speaking to.