Optimizing Text Correction For Voice Based IoT Smart Building Virtual Assistants

Abstract

The integration of Virtual Assistants (VAs) within Smart Building Internet of Things (IoT) ecosystems is increasingly critical, particularly for interpreting user commands via Automatic Speech Recognition (ASR). This paper presents an in-depth performance analysis of text correction algorithms on a Raspberry Pi 4—a cost-effective and widely used computing solution in smart building applications. Due to the absence of GPU acceleration for Python on ARM architecture, a specialized dataset was developed to benchmark algorithmic performance, focusing on correction times and accuracy. Our study utilized a near-real-world experimental setup, deploying Docker containers to simulate IoT MQTT brokers, a Smart Building Platform, and Rasa for dialogue management. Among the algorithms tested—Edit distance, Jaccard, FuzzPartialRatio, FuzzSortRatio, MLE, and Norvig Spell—the Edit distance and Norvig Spell emerged as leaders in accuracy, achieving an 84% success rate in text correction. Notably, the Edit distance algorithm demonstrated superior speed, vital for real-time processing demands. The Fuzz Sort Ratio algorithm distinguished itself with the fastest correction time at 31.6 milliseconds, albeit with a slight compromise on accuracy, attaining a 79% success rate. Consequently, the Edit distance algorithm is recommended for applications where accuracy and response time are paramount, while the Fuzz Sort Ratio is preferable for scenarios where speed is the overriding priority. This research paves the way for future exploration into the computational impacts of these algorithms and the exploration of neural network-based methods to further enhance text correction capabilities in smart building automation systems.