Research and Development of Intelligent Electrical Appliance Control using Deep Learning Speech Recognition

Yap, Xin Wei (2020) Research and Development of Intelligent Electrical Appliance Control using Deep Learning Speech Recognition. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text Yap Xin Wei.pdf Restricted to Registered users only Download (6MB)

Abstract

Most of the electrical appliances in our homes are controlled and operated manually through switches located at the wall. However, this could bring difficulties to people in some cases, such as for handicap and visual impaired people, or when the switches are difficult to be reached. Speech recognition has been a constant developing technology nowadays as it brings convenience to our daily lives. In this project, researches has been conducted to compare some popular speech recognition technologies which include client-server speech recognition system, hidden Markov model (HMM) speech recognition system and artificial neural network (ANN) speech recognition system. The purpose of this project is to develop an intelligent electrical appliance control system using deep learning speech recognition technique and convolutional neural network (CNN) is used as the architecture of the network. A dependent speech recognition (DSR) system and an independent speech recognition (ISR) system have been designed and the accuracy of each network is tested by same sets of validation and testing data. The testing data are obtained from different conditions, which include speech recorded in a clean environment, speech recorded in a clean environment and 15 cm away from the microphone, speech recorded in a clean environment and 30 cm away from the microphone, speech recorded in a dirty environment, speech recorded in a dirty environment and 15 cm away from the microphone, and speech recorded in a dirty environment and 30 cm away from the microphone. For the DSR system, the neural network has 21x1 layers and 12 convolutional filters. It achieves a validation accuracy of 100 %, training error of 0.76453 % and average testing error of 14.8472 %. For the ISR system, the neural network has 33x1 layers and 12 convolutional filters. It achieves a validation accuracy of 94.81 %, training error of 0 % and average testing error of 20.524 %. After the network is completely trained, speech can be inputted through a microphone and it will be classified by the network. Total of 36 commands in English, Mandarin and Malay can be detected by the trained system. Once a command is identified, Arduino Uno will control the operation of electrical appliance such as light, fan, aircond and TV based on the command received.