IEEE Project Abstract

Speech-based interfaces are convenient and intuitive,and therefore, strongly preferred by Internet of Things (IoT)devices for human-computer interaction. Pre-defined keywordsare typically used as a trigger to notify devices for inputting thesubsequent voice commands. Keyword spotting techniques usedas voice trigger mechanisms, typically model the target keywordvia triphone models and non-keywords through single-state fillermodels. Recently, deep neural networks (DNN) have shown betterperformance compared to hidden Markov models (HMM) withGaussian mixture models (GMM), in various tasks includingspeech recognition. However, conventional DNN-based keywordspotting methods cannot change the target keywords easily, whichis an essential feature for speech-based IoT device interface.Additionally, the increase in computational requirementsinterferes with the use of complex filler models in DNN-basedkeyword spotting systems, which diminishes the accuracy of suchsystems. In this paper, we propose a novel DNN-based keywordspotting system that alters the keyword on the fly and utilizestriphone and monophone acoustic models in an effort to reducecomputational complexity and increase generalizationperformance. The experimental results using the FFMTIMITcorpus show that the error rate of the proposed method wasreduced by 36.6%.