David L. Bick

Hi! My name is David Bick and I work in Applied ML at Cerebras Systems. Cerebras is the leading startup working on hardware acceleators for deep learning.

Previously I spent six years at Carnegie Mellon University. I first earned a Bachelors in Statistics and Machine Learning, and then published three papers under Prof. Bhiksha Raj during a Research Masters in the Language Technologies Institute, within the School of Computer Science.

My work has been in natural language processing and speech processing. My research followed a line of work on improving the perceptual quality of machine-generated speech. Now I work in NLP with LLMs, and have worked on projects at CMU and Cerebras that include multi-modal QA, instruction-fine-tuning, etc.

Research Details:

To improve perceptual quality, we created deep learning estimators of fine-grained speech characteristics, whose importance we identified from the acoustic-phonetics literature. A useful analogy is that they are a “fingerprint” of natural speech. Our work has focused on different ways to optimize networks to produce speech that has this same “fingerprint”.

These characteristics were previously calculated using traditional non-differentiable signal processing techniques. Our neural estimators are differentiable and thus open up their application in optimizing other networks. In the last year we have presented this work in 3 papers across 2 conferences, and at an invited talk to Meta Reality Labs Research Audio.