Ashutosh Chaubey

I am a CS PhD student at the Institute for Creative Technologies, University of Southern California, where I am advised by Prof. Mohammad Soleymani at the Intelligent Human Perception Lab. I am a Bronze Medallist from the 2021 batch of Indian Institute of Technology, Roorkee

Prior to this, I was a Founding Research Engineer at Anoki AI where I worked on projects related to content retrieval using audio/text and text-to-speech. I have also worked at LG Ad Solutions as a Data Scientist, where I spent some time working on speaker recognition, automatic content recognition using audio and voice cloning. Over the past I have interned at Adobe Research, where I worked with Dr. Sumit Shekhar on active learning for content labelling in documents. I have also interned at Video Analytics Lab, IISc. Bengaluru where I worked with Prof. R. Venkatesh Babu on human pose estimation from a single RGB image. Back at my undergraduate college, I worked with Prof. R. Balasubramanian on automatic evaluation of machine synthesized speech.

I am always eager to collaborate on research with people in academia as well as industry. Please reach out to achaubey at usc dot edu to discuss potential collaborations.
Currently, I am looking for Research/Applied Scientist internship positions for Summer 2025. Please reach out if you have open positions.

Email  /  CV  /  Google Scholar  /  Linkedin  /  Github

I have experience working on speaker recognition and speech processing systems. More recently, I am working on computer vision where I deal with problems related to multi-person interaction, behaviour generation and diffusion models.

Proposed two novel approaches for imposter identification in unseen speaker recognition. One of the approaches is speaker-specific thresholding and the other relies on meta-learning to decouple the problem of imposter identification from speaker identification.

Inspired from their use in computer vision, we use relation networks for the task of speaker recognition and propose enhancements in terms of global supervision and faster training regime.

We propose a reinforcement policy based active learning approach for document content labelling tasks such as object detection, layout detection and named entity recognition.

We present a comprehensive survey on universal adversarial perturbations, both attacks and defenses along with future directions for the topic.

We propose a technique which leverages the discriminator from a GAN based TTS model for automatic evaluation of machine synthesized speech.

Presented my paper A Generative Adversarial Network Based Ensemble Technique for Automatic Evaluation of Machine Synthesized Speech at the Student Academic Conference at Inter IIT Tech Meet 2019.

My hobbies include cooking and playing guitar. I am obsessed with the songs of Prateek Kuhad and Anuv Jain.