Pavan Seshadri

I'm a second-year masters student at Georgia Tech advised by Dr. Alexander Lerch in the Music Informatics Group.

I am broadly interested in topics spanning speech/audio/music and language representation learning, including information retrieval, recommendation systems, and multimodal learning, especially for systems that facilitate artistic creativity or discovery.

My current work has centered around topics in music information retrieval (MIR), soundscape detection, and music recommender systems.

Prior to starting graduate school, I was a Machine Learning Engineer at Amazon where I worked on NLP research and infrastructure for product classification.

I recieved my bachelors degree in Computer Science from Georgia Tech in 2021, with a minor in Music Technology. My undergraduate research focused on representation learning methods for music performance assessment.

I will be available and looking for full-time roles starting December 2024.

Email  /  CV  /  Linkedin  /  Github  /  Google Scholar

We propose a contrastive learning task to model temporal negative feedback within the objective function of a sequential music recommender, and demonstrate performance gains over feedback-agnostic systems.

We introduce the new audio analysis task of pedestrian detection and present a new large-scale dataset for this task. While the preliminary results prove the viability of using audio approaches for pedestrian detection, they also show that this challenging task cannot be easily solved with standard approaches.

Contrastive loss based neural networks are able to exceed SoTA performance for music performance assessment (MPA) regression tasks by learning a better clustered latent space.

We present a study using self-attentive architectures for next-track sequential music recommendation. We additionally propose a contrastive learning subtask to learn session-level track preference from implicit user signals, resulting in a 3-9% top-K hit rate performance increase relative to baseline negative feedback-agnostic approaches.

We propose a dataset, AVASpeech-SMAD, which provides frame-level music labels for the existing AVASpeech dataset, originally consisting of 45 hours of audio and speech activity labels.