Detecting Persian speaker-independent voice commands based on LSTM and ontology in communicating with the smart home appliances

Abstract

Nowadays, various interfaces are used to control smart home appliances. The human and smart home appliances interaction may be based on input devices such as a mouse, keyboard, microphone, or webcam. The interaction between humans and machines can be established via speech using a microphone as one of the input modes. The Speech-based human and machine interaction is a more natural way of communication in comparison to other types of interfaces. Existing speech-based interfaces in the smart home domain suffer from some problems such as limiting the users to use a fixed set of pre-defined commands, not supporting indirect commands, requiring a large training set, or depending on some specific speakers. To solve these challenges, we proposed several approaches in this paper. We exploited ontology as a knowledge base to support indirect commands and remove user restrictions on expressing a specific set of commands. Moreover, Long Short-Term Memory (LSTM) has been exploited for detecting spoken commands more accurately. Additionally, due to the lack of Persian voice commands for interacting with smart home appliances, a dataset of speakerindependent Persian voice commands for communicating with TV, media player, and lighting system has been designed, recorded, and evaluated in this research. The experimental results show that the LSTM-based voice command detection system performed almost 1.5% and 13% more accurately than the Hidden Markov Model (HMM)-based one, in scenarios 'with' and 'without ontology', respectively. Furthermore, using ontology in the LSTM-based method has improved the system performance by about 40%.