publications
publications by categories in reversed chronological order. generated by jekyll-scholar.
2025
- Super-Resolution via Learned PredictorSampath Kumar Dondapati, Omkar Nitsure, and Satish MulletiIn 2025 National Conference on Communications (NCC), 2025
Frequency estimation from measurements corrupted by noise is a fundamental challenge across numerous engineering and scientific fields. Among the pivotal factors shaping the resolution capacity of any frequency estimation technique are noise levels and measurement count. Often constrained by practical limitations, the number of measurements tends to be limited. This work introduces a learning-driven approach focused on predicting forthcoming measurements based on available samples. Subsequently, we demonstrate that we can attain highresolution frequency estimates by combining provided and predicted measurements. In particular, our findings indicate thatusing just one-third of the total measurements, the method achieves a performance akin to that obtained with the complete set. Unlike existing learning-based frequency estimators, our approach’s output retains full interpretability. This work holds promise for developing energy-efficient systems with reduced sampling requirements, which will benefit various applications.
- Learnable Kernels for FRI – Joint Kernel Encoder Optimization and Hardware ValidationOmkar Nitsure, Sampath Kumar Dondapati, and Satish Mulleti2025
Finite Rate of Innovation (FRI) sampling techniques provide efficient frameworks for reconstructing signals with inherent sparsity at rates below Nyquist. However, traditional FRI reconstruction methods rely heavily on pre-defined kernels, often limiting hardware implementation and reconstruction accuracy under noisy conditions. In this paper, we propose a robust, flexible, and practically implementable framework for FRI reconstruction by introducing novel learnable kernel strategies. First, we demonstrate effective reconstruction using known, fixed kernels such as truncated Gaussian and Gaussian pair kernels, which mitigate the requirement that the samples should have a sum-of-exponentials (SoE) form. Next, we extend this concept by jointly optimizing both the sampling kernel and reconstruction encoder through a unified learning approach, yielding adaptive kernels that significantly outperform traditional methods in resolution and noise robustness, with reduced sampling rates. Furthermore, we propose a practical hardware realization by representing kernels as sums of two exponential decay signals with jointly optimized poles, facilitating compact, efficient analog implementations. Our approach is validated experimentally through hardware implementations using a unity-gain Sallen-Key analog filter, achieving accurate real-world signal recovery. The developed convolutional neural network-based encoder substantially reduces computational complexity, demonstrating competitive performance with fewer parameters, making our method particularly suitable for resource-constrained, edge-based deployments.
2024
- Improving Self-supervised Pre-training using Accent-Specific CodebooksDarshan Prabhu, Abhishek Gupta, Omkar Nitsure, and 2 more authorsIn Interspeech 2024, 2024
Speech accents present a serious challenge to the performance of state-of-the-art end-to-end Automatic Speech Recognition (ASR) systems. Even with self-supervised learning and pre-training of ASR models, accent invariance is seldom achieved. In this work, we propose an accent-aware adaptation technique for self-supervised learning that introduces a trainable set of accent-specific codebooks to the self-supervised architecture. These learnable codebooks enable the model to capture accent specific information during pre-training, that is further refined during ASR finetuning. On the Mozilla Common Voice dataset, our proposed approach outperforms all other accent-adaptation approaches on both seen and unseen English accents, with up to 9% relative reduction in word error rate (WER).