We want what you want. Companions you actually want to talk to. Game characters that feel real. Virtual actors where the screen is the game engine. Robots that can really do tasks, not staged demos.
Meet our avatar. An embodied mind, learning the world the way you did. Livestream coming soon.
Fig. 01Today's AI sounds smart but makes things up. It has no sense of space or time. It has never touched the world.
The only path to real intelligence is an AI that learns the way humans do, through a body that sees, touches, hears, talks, sings, moves, feels gravity, experiences time, feels pain and pleasure, and learns the difference between self and world.
This needs three things. A body, a mind that can generalize from experience, and a world. Today's AI cannot generalize. It does not understand what it is doing. Not its body. Not the space around it. Not the consequences of acting. Real generalization needs all of these understood together. That is why every robotics demo is staged.
Picture asking a robot, in a room it has never seen, to walk around two tables you point at and bring back the keys off the third. Knowing they are yours because you threw them. No public demo can do this. Not in simulation. Not on hardware. That is the wall every robotics company hits.
We are building all three. The body. The mind. The world. Our patented architecture (US 11,537,850 B2 + 6 pending patents) describes how movement, touch, pain, and pleasure together shape the way an embodied AI perceives the world and develops intent. The same way a developing infant learns but a lot faster.
High fidelity deforming touch.
Fig. 02A voice built from physics, not data.
Most synthetic voices are a remix of someone who was once recorded. BioFold models the human vocal tract itself, and our avatar learns to control it the way a child learns to speak. The result is a voice that runs in real time, sounds identical every time you call it, and can belong to a person who has never existed.
Physically modeled voice is a small and rarely explored field. BioFold is still early, but we believe it may already be the most human sounding in the world.
None of what you’re about to hear was generated by AI. Every sound comes straight from the physical model, tuned by hand. Nothing is sampled or stitched from a recording. The next step is handing the controls to our AI, which will learn to drive the model the way a child learns to use their voice, and take it far beyond anything hand-tuning can reach.
One clip is a real human singer. The other two are BioFold rebuilding that same voice from a physical model of the vocal tract. No recording of this person is used to make them. Pick a voice and a vowel, then play them back to back.
Voice
Vowel
A real voice never holds a note perfectly steady. Its loudness shimmers slightly, moment to moment. That flutter is much of what separates “alive” from “machine.” Here is the same synthesized note, before and after BioFold adds it.
Voice
Vowel
Real recordings are from VocalSet (J. Wilkins, P. Seetharaman, A. Wahl, B. Pardo, “VocalSet: A Singing Voice Dataset,” ISMIR 2018, Northwestern University), licensed CC BY 4.0, source doi.org/10.5281/zenodo.1193957. Real clips are excerpts (silence trimmed). The reproduction clips are our own synthesized reproductions derived from parameters extracted from those recordings (© Mind Machine Learning); not part of or endorsed by VocalSet.