Collaborations with Adobe Research and Qualcomm Technologies selected to receive Dean for Research Innovation funding

Written by
Catherine Zandonella, Office of the Dean for Research
April 18, 2019

Two projects that bring together faculty researchers and industry to transform discoveries into real-world technological applications have been awarded support from the Dean for Research Innovation Fund for Industrial Collaborations.

The fund was established by the Office of the Dean for Research in recognition of the essential role that industry often plays in bringing research innovations to fruition and making them available to benefit society. This year’s winning proposals were selected by a faculty review panel for their quality, originality and potential impact.

The winning projects are:

Deep learning for natural-sounding speech synthesis

Adam Finkelstein and Zeyu Jin
Adam Finkelstein and Zeyu Jin, a collaborator at Adobe Research. Photo by Mark Czajkowski

A collaboration between Princeton researchers and Adobe Research — the division of Adobe that shapes early-stage ideas into innovative technologies — seeks to develop advances in deep learning for understanding human dialogue and generating natural-sounding speech, and ultimately enriching human-computer interactions. Deep learning is a type of machine learning in which computer systems solve complex problems by analyzing huge datasets. Already, advances in deep learning have streamlined communication with virtual voice assistants such as Siri and Alexa. However, generating the natural-sounding synthetic speech patterns of a specific person remains challenging. To develop this capability, Adam Finkelstein, professor of computer science, and his team will collaborate with Adobe Research scientists Gautham Mysore, Zeyu Jin and Richard Zhang to improve numerical measures, called loss functions, that indicate how well a machine has learned from training data – a crucial component of deep learning. The team will design loss functions for speech synthesis based on data gathered from human volunteers asked to judge how natural an audio clip sounds, and compare how similar two audio clips sound. To further improve efficiency and scale, the team will devise ways to enable the machine to learn from human responses in real-time. The resulting datasets aim to improve speech processing and synthesis, and may lead to new applications such as making cellphone-recorded speech sound like a studio recording, and the ability to edit a recorded narration simply by typing text.

Opening up the future of wireless communications

Researchers standing
Kaushik Sengupa, assistant professor of electrical engineering, left, with graduate student Chandrakanth Reddy Chappidi, postdoctoral research associate Tushar Sharma, and graduate student Zheng Liu. Image courtesy of the researchers

A collaboration between Qualcomm Technologies, Inc. and researchers at Princeton aims to open up new capacity for mobile phones and other wireless devices to connect to the internet by expanding use of the electromagnetic spectrum, the "highway" that carries wireless signals. Qualcomm intends to support Princeton researchers who seek to develop circuits capable of harnessing sections of the spectrum difficult to access with current electronics. Kaushik Sengupta, assistant professor of electrical engineering, will coordinate with Qualcomm researchers to design technology for devices that harness the millimeter wave (mm-wave) frequency spectrum. To make the most of these channels, instead of operating at a fixed frequency like a car stuck in the same lane of traffic, the devices can rapidly switch frequencies like a vehicle weaving in and out of lanes to avoid traffic. With this dynamic use of the spectrum, researchers can build programmable wireless transmitters that allow faster and more flexible communication between the many devices – from autonomous vehicles to health monitors – expected to come online in the near future.