RNN, LSTM, and Voice Signal Model for Electrical Switch Regulator

The creation of platforms for systematic control of electric power switches based on the hands-on application of Artificial Intelligence in day-to-day life lessens the prospect of unintended switch initiation. It can enhance security by ensuring that only authorized users receive responses. Physically challenged individuals also need systems bereft of point-point contacts or interactions with electrical or power switches. Some of the known methods for achieving these objectives include smart objects, the Internet of Things, and biometric technologies, each with its strengths and weaknesses. This paper proposed a voice signal model for remote control of electrical switches. The model uses a voice recorder connected to an Arduino microcontroller to boost the audio or voice signal from the user, while a voice sensor is also linked to a power switch relay to acquire the voice signal for registration, training, verification, and processing. The Arduino microcontroller sensor runs TinyML and TensorFlow Lite environment sensors while operating at an adjustable voltage. A switch relay was required for limiting the voltage to a required level based on synergy with the Arduino microcontrollers. A Wi-Fi module was also used for launching the microcontroller and the TCP/IP connections based on Hayes-style commands. The system runs with an electromechanical device designed for the flow of electric current to open or close the electrical circuit. The user voice recognition leverages Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) networks to guarantee effective capturing of temporal dependencies in sequential data typical of audio signals. The implementation of the framework on specialized hardware and software has established its ability to effectively classify spoken commands as either ON or OFF independent of the speaker’s identity, based on comparison of the input features. Performance evaluation based on the confusion matrix established that a very high percentage of the input commands were correctly recognized with an accuracy of 90%.