A small lab with one question.
Vyakti takes its name from the Sanskrit word for a person: the individual, that which is made manifest. The question is whether a machine can sustain identity, history and presence without pretending to be human.
How we work.
Behaviour before benchmarks
The score exists to describe the behaviour. When the two disagree, we believe the conversation and rewrite the metric.
Say what does not work
Every system has a range where it works reliably and a range where it does not. Reporting only the first obscures the work still required.
Research through one product
Research that never reaches real use stays theoretical. Meera is where the work has to survive contact with time, context and shared history.
Culture is not a setting
There is no neutral way to have a conversation. Building as though there is produces a system that is subtly wrong for everyone.
Careers
People we hope to meet.
If your work overlaps with these areas, write and tell us which part of the problem you would take on.
Speech and voice
Speech-to-speech modelling, prosody and expressive synthesis. You have shipped audio models and care about how they sound to a person, not only how they score.
Conversation systems
Turn-taking, endpointing and full-duplex dialogue. Comfortable with real-time constraints and with linguistics literature that predates deep learning.
Long-horizon evaluation
Designing the studies that tell us whether any of this is working. Experimental design, human subjects, and a healthy suspicion of your own results.
Design engineering
The surfaces people meet these systems through. Interface, motion and sound, built to the standard the research is held to.