echosonic

However a ubiquitous way of audio signal recognition, machine learning based technology suffers latency and privacy and security issues due to enormous computational load and huge amount of data transfer between the cloud and sensors. echosonic employs unique non-linear physics commonly existing in MEMS microphone and analog circuit to have machine learning directly operated on the microphones. As such, the audio signal processing could be done in a way with less than 1% of the computational power and 5% of the required data for software upgrading and potentially provides us a new way of constructing deep neural networks on the IoT. I'm the technical lead of the project, spearheading the software and hardware development. My job responsibilities include identifying technology advisory board members, codevelopment of new technologies through university collaborations, building engineering team, new IP ideation, and constructing prototypes.

echosonic device is capable of converting audio signal directly to features (as shown in the following figure) for machine perception using the sensing mechanisms of sensors. As such, it bridges the gap between the data collecting and processing, enables a more consolidated and computational-power friendly type of audio signal processing. More importantly, the direct feature engineering unifies the audio signal with visual signal enabling a shared processing approach using machine learning with different types of data.