Manvi Agarwal

I am a PhD student working with Gaël Richard and Changhong Wang at Télécom Paris, Institut Polytechnique de Paris, supported by the ERC-funded Hi-Audio project.

I am interested in using Machine Learning methods for analyzing and extracting insights from temporal sequence data. My particular testbed is musical signals. My PhD research looks at how inductive biases can be introduced into Transformers by using the positional encoding module to insert musically-relevant structural information.

Before starting my PhD, I worked as a Machine Learning Engineer at Sentinels in Groningen, The Netherlands, building solutions for the financial transaction processing industry. I completed my Masters in Artificial Intelligence with cum laude distinction at the University of Groningen in 2020. I completed my Bachelor of Arts in History at Presidency University. I spoke to The Telegraph India about my journey from the humanities to the sciences here.

The best way to reach me is to write me an email at firstname dot lastname at telecom-paris dot fr.

publications

ICASSP

Structure-Informed Positional Encoding for Music Generation

Manvi Agarwal, Changhong Wang, and Gaël Richard

In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2024

Abs arXiv Code Website

Music generated by deep learning methods often suffers from a lack of coherence and long-term organization. Yet, multi-scale hierarchical structure is a distinctive feature of music signals. To leverage this information, we propose a structure-informed positional encoding framework for music generation with Transformers. We design three variants in terms of absolute, relative and non-stationary positional information. We comprehensively test them on two symbolic music generation tasks: next-timestep prediction and accompaniment generation. As a comparison, we choose multiple baselines from the literature and demonstrate the merits of our methods using several musically-motivated evaluation metrics. In particular, our methods improve the melodic and structural consistency of the generated pieces.