Files
Abstract
The emergence of voice-controlled systems has transformed the way users engage with technology, providing unparalleled ease and accessibility. Nonetheless, these systems encounter difficulties in distinguishing intended commands from unintentional triggers, especially when competing with TV broadcast sounds. This dissertation addresses this issue by creating and testing multiple techniques to differentiate ambient sound environments from TV broadcast audio. The study presents four unique methods for detecting TV audio: Energy Balance Metric, YAMNet-based TV Audio Detection, Fine-Tuning YAMNet with LSTM Backend, and an Ensemble of YAMNet-LSTM Models. Each method employs audio signal processing and machine learning techniques to provide a comprehensive and diverse perspective on TV sound detection. Comprehensive evaluations are conducted using various datasets, including EAR-Aging, EAR-Divorce, SINS, and a custom lab dataset. These datasets cover a broad range of real-world situations, including natural ambient recordings and controlled laboratory environments, providing a comprehensive assessment of each technique. The assessment centers on crucial benchmarks such as the ability to detect TV noise, identifying non-TV sounds with precision, and overall performance under different acoustic conditions and hardware setups. The thesis findings highlight the capabilities and drawbacks of each technique for precisely identifying sounds from TVs, providing valuable insights on their practical use in voice-controlled systems. Particularly, the Ensemble model offers a promising combination of unique features, resulting in improved accuracy and adaptability. This study contributes innovatively to the ambient sound classification domain, presenting novel solutions to accurately identify TV sounds. The optimization of voicecontrolled systems can enhance reliability and improve user experience in everyday environments, opening new avenues for exploration. Future work can extend these methodologies to broader applications, including smart home automation and assistive technologies for people with disabilities.