Mason L. Wang

I am an EECS PhD student at MIT CSAIL, where I am working with Professor Anna Huang. My research is at the intersection of audio, machine learning, and signal processing.

I received my master's in Electrical Engineering at Stanford University working at the Stanford Vision and Learning Lab. I was advised by Jiajun Wu and Mert Pilanci.

You can contact me at ycda [at] stanford [dot] edu. Or, you can find me on Twitter, LinkedIn.

Research Interests

Capturing Real Auditory Scenes: Recently, my work has been on virtualizing real auditory scenes and acoustic spaces. For instance, imagine being able to capture a video of a concert, and then moving around the concert space freely. Imagine taking several videos of a fireworks show, and compiling them into an interactive 3D experience. Perhaps you could capture the intrinsic acoustic properties of your living room, in a way that allows you to listen to your favorite artist there.

Differentiable and Inverse Audio Rendering: Audio renderers often require slow and non-differentiable techniques. This makes it difficult to fit to real scenes via gradient-based optimization processes, and thus, often results in audio simulations that are not accurate to the real-world sounds they attempt to replicate. Inspired by visual inverse rendering and capture techniques, I believe combining physical inductive biases with machine learning can help us fit simulations to real scenes, and thus make them more accurate.

AI assisted Sound Design and Music-making: Making music requires many steps: writing melodies/themes, chord progressions, arrangement, sound design, mixing, mastering, etc. Some musicians are more inclined towards certain parts of this process. My goal is to provide musical artists with controllable assistance for parts of the music-making process they are unfamiliar with.

Publications

We create a method of capturing real acoustic spaces from 12 RIR measurements, letting us play any audio signal in the room and listen from any location/orientation. We develop an 'audio inverse-rendering framework' that allows us to synthesize the room's acoustics (monoaural and binaural RIRs) at novel locations and create immersive auditory experiences (simulating music).

Humans induce subtle changes to the room's acoustic properties. We can observe these changes (explicitly via RIR measurement, or by playing and recording music in the room) and determine a person's location, presence, and identity.

Everyday objects possess distinct sonic characteristics determined by their shape and material. RealImpact is the largest dataset of object impact sounds to date, with 150,000 recordings of impact sounds from 50 objects of varying shape and material.

We use a dataset of full-mix songs, stem-subtracted songs, and LLM-generated edit instructions to train a stem editing/insertion diffusion model.

Education and Experience

News

12/20/24

04/13/24

02/26/24

09/21/23

02/27/23

Subtractive Training is accepted to ICASSP 2025!

First-author submission to ISMIR 2024!

Hearing Anything Anywhere is accepted to CVPR 2024!

SoundCam is accepted to NeurIPS Datasets and Benchmarks 2023!

RealImpact is accepted to CVPR 2023!

Music Works

Knowledge Map

Website Template based on Jon Barron's.