Rajalaxmi Rajagopalan

I’m a Ph.D. student in the Electrical and Computer Engineering Department at the University of Illinois, Urbana Champaign (UIUC).

I am working with Dr. Romit Roy Choudhury in the Signals and Inference Research Group (SiNRG)

Research Interests

My research interest is in Black-box optimization problems and other problems in signal processing and acoustics. My current research is in audio personalization in earables and universal speech enhancement.

Current Research Projects

  • Sample Constrained Black-Box Optimization for Personalization: We consider the problem of personalizing content to a user’s taste. Content could be audio signals in a hearing aid, customized images, etc. Given the content, we intend to adjust it with a linear filter h. Our goal is to find the optimal filter h that will maximize the user’s personal satisfaction f(h).

    Finding h is difficult as the function f(h) is unknown. Optimizing f(h) using the human in the loop, while constrained by the number of samples is a sample-efficient black-box optimization problem. We build on the Bayesian Optimization framework. We achieve sample efficiency by developing kernel learning techniques that learn a unique kernel that models function structure. Our techniques are extended to other black-box problems in many areas beyond personalization.

  • Self-supervised Universal Speech Enhancement: Speech is often corrupted by many distortions like ambient noise, reverberation, limited bandwidth, codec artifacts, interfering speakers etc. In practical scenarios, there is lack of high-quality clean speech references. Thus, we develop a universal speech enhancer that (1) can tackle a broad range of distortions simultaneously, (2) trained without access to large-scale clean speech corpus, (4) trained in a self-supervised manner, and (3) offers robust performance for enhancement downstream tasks.

    We use a pre-training-finetuning framework. A Masked Spectrogram Autoencoder based on Vision transformers is the model of choice. The pre-trained embeddings are then used by fine-tuning models trained on a small amount of paired data for specific downstream tasks like denoising, dereverberation, source separation, and bandwidth extension.
  • CV

    Here’s my CV.