Towards Development of Smart and Reliable Voice Based Personal Assistants

Abstract

Due to advent of Machine Learning and Internet of Things, communication between humans and machines has increased substantially. This leads to need for efficient software that can enable machine-to-human interaction seamlessly. Internet of Things leads to automation as it is aimed at frequent data collection without continuous human intervention. Moreover, there are many tasks which can easily be delegated to automated software programs (i.e. agents) so as to provide flexibility and ubiquitous service to users. This, in turn, requires development of semantic-enabled, intelligent agents that can also capture implicit context-specific details from the user queries on World Wide Web (WWW). Querying the WWW using voice commands is the current trend. However, there are some issues. In this work, we focus on building smart and reliable agents to overcome a big challenge faced by voice based personal assistants. We have attempted to address security issues in voice biometrics systems. These systems are most susceptible to playback spoofing attacks due to availability of smartphones to any end user. We have used ASVspoof 2017 dataset for implementation and presented our findings for replay spoofing detection using fusion of short-term spectral features. Results using inverse MFCC features are promising and point to directions for further research with use of these potential features.