Madhushankara Maila

Hydroacoustic signature-based ship detection (HSD) involves identifying and classifying vessels by analyzing the underwater sound patterns they emit, including propeller noise, engine vibrations, and hull flow disturbances. This technique is useful in naval surveillance, anti-submarine warfare, maritime traffic monitoring, and underwater environmental studies where visual or radar detection is limited. Since HSD is independent of visibility conditions and covert due to its passiveness, it can be used as a complementary modality along with traditional ship detection methods. However, machine-learning based automatic HSD is sensitive to factors like variability in ocean acoustic environments, background noise, similarity of signatures among different classes of ships, etc., especially in low-resource conditions where there is lack of sufficient training data to train a generalized HSD model. Furthermore, out-of-domain (OOD) detection is necessary for reliable surveillance in real-world applications. Motivated by all these, in this paper, we present different methods to improve the performance of low-resource HSD with OOD detection capability. Specifically, we first propose to use pre-trained PaSST and EAT models for efficient feature extraction for low-resource HSD. Since different encoder layers of these pre-trained models encode the input differently, a detailed layer-wise analysis of the performance is conducted on both models to select the optimum layer for HSD-specific feature extraction. Following this, we incorporate OOD detection capability to the HSD model using GMM-based, K Nearest Neighbor (KNN) based and OpenMax-based thresholding approaches. The results obtained on a publicly available dataset indicate that the HSD model trained using PaSST features, when coupled with the GMM-based OOD detector achieves the best performance, with an AUROC of around 93% and an accuracy score of nearly 86%.

Infants use cry as their primary mode of communication, which they use to signal hunger, discomfort, illness etc., to their caregivers. Understanding and interpreting these cries can provide critical insights into their well-being and aid in timely intervention. However, manual analysis of infant cry is subjective and prone to variability, especially when parents/caregivers have limited experience. While deep learning (DL) based infant cry classification (ICC) systems can be handy in such situations, they need large datasets to train robust models, without which they become vulnerable to challenges like variability in cry patterns across infants, noise in real-world recordings, etc. One way to improve the performance in such low-resource conditions is to use transfer-learning, where features extracted from pre-trained model are used to build models for the task-at-hand. Motivated by this, we propose to build ICC system using features from pre-trained wav2vec 2.0 and whisper models and compare their performance with the traditional ICC system that uses MFCC features. Since different layers of wav2vec 2.0 and whisper might encode the required information differently, we first empirically determine their relevance to ICC. In order to best utilize the complementary information that may be present in the features extracted from different layers and time-steps, we then propose to fuse them together. We propose to use different fusion strategies such as statistical pooling, attentive statistical pooling, linear-merge and attention-merge based approaches to aggregate the features. Furthermore, we explore different deep-learning architectures such as LSTM and x-vector-based architecture to perform cry classification. Obtained results indicate that attention-merge on wav2vec 2.0 large features combined with x-vector gives the best performance.

An electrolarynx, a widely used assistive device for voice restoration for individuals with laryngeal impairments, often suffers from mechanical buzz and poor intelligibility, limiting its everyday applicability. This work presents a modified recursive least squares (M-RLS) algorithm specifically designed to enhance electrolarynx speech by attenuating the excitation noise characteristic while preserving speech quality. The proposed work is evaluated on data across 44 speech samples and benchmarked against a baseline method using three objective metrics: the perceptual evaluation of speech quality (PESQ), short-time objective intelligibility (STOI), and the signal-to-noise ratio (SNR). The proposed M-RLS approach demonstrated notable improvements, with PESQ scores reaching 3.01, a median STOI value of 0.7, and an SNR performance comparable to that of the baseline. In addition, to evaluate its feasibility in embedded systems, the algorithm was synthesized in 15 nm technology to form an application-specific integrated circuit (ASIC), confirming its suitability for real-time, low-power speech enhancement applications. The design occupies 0.289 mm², operates at 1 GHz, and consumes only 16.9 mW, yielding a power-delay product of 15.41 pJ, making it well suited for energy-constrained environments such as electrolarynx-based speech enhancement systems. The proposed method is also quick prototyped via a DE10 field programmable gate array (FPGA) with a maximum operating frequency of 253.8 MHz.

Sustainability in in-vehicle communication comprises the integration of wireless technology for faster and more reliable connectivity. Driverless vehicles need to schedule various tasks on a priority basis. The first-in-first-out (FIFO) data structure is critical for handling sensor data and communication information, and a memory queue is important for controlling data flow through organized read and write processes. This study focuses on the design, verification and synthesis of a FIFO with a size of 4k bytes. By operating on a single clock signal, synchronous FIFO enables smooth data transmission between a source and destination within the same clock domain. The designed FIFO resulted in an area of 41.401[ ]mm², a power dissipation of 0.506[ ]mW and a critical path delay of 0.03[ ]ps when a 15[ ]nm open-cell library from Nangate was used. The design is also implemented using an Arria II field programmable gate array (FPGA), which results in a maximum clock frequency of 33.33[ ]GHz.