Wake Word Accuracy in Amazon Alexa Devices
Academic Level at Time of Presentation
Senior
Major
Computer Information Systems
List all Project Mentors & Advisor(s)
Dr. Marcia Combs-Ford
Presentation Format
Oral Presentation
Abstract/Description
Virtual digital assistant software is becoming more and more popular in every new piece of technology, from smart home devices to appliances. One of the most notable creations, and frontrunners in the industry, is Amazon Alexa. Millions of Alexa devices have been sold and their success continues to grow, with expected sales of Alexa devices to exceed $10 billion by 2020. However, there are growing privacy concerns among users about their virtual digital assistant security and what these devices are actually recording. Virtual digital assistants begin recording a user’s voice commands when the “wake” word is spoken. Alexa devices can be set to awaken to different words: Amazon, Alexa, or Computer. Inadvertent recordings are made when the device hears the wake word in normal conversation rather than a command, or it can interpret TV or background noise as a command and begin recording. These recordings can cause a backlog in the history file and decrease the accuracy in wake word identification. This study will analyze two Amazon Alexa voice recording datasets collected over a period of 45 days in an attempt to answer the research question: “which wake word yields the fewest inadvertent recordings.” The experimental network will consist of two Amazon Echo Dot’s, two Echo Show’s, and one Echo Tall, with all devices equipped with the Alexa voice software. Based on my findings, I will provide recommendations for wake word safe selection operation and possible amendments to future software updates as well as descriptive analysis of other voice recording characteristics such as of dialect recognition, tone of voice, and frequency of recordings.
Fall Scholars Week 2019 Event
Honors College Senior Theses
Wake Word Accuracy in Amazon Alexa Devices
Virtual digital assistant software is becoming more and more popular in every new piece of technology, from smart home devices to appliances. One of the most notable creations, and frontrunners in the industry, is Amazon Alexa. Millions of Alexa devices have been sold and their success continues to grow, with expected sales of Alexa devices to exceed $10 billion by 2020. However, there are growing privacy concerns among users about their virtual digital assistant security and what these devices are actually recording. Virtual digital assistants begin recording a user’s voice commands when the “wake” word is spoken. Alexa devices can be set to awaken to different words: Amazon, Alexa, or Computer. Inadvertent recordings are made when the device hears the wake word in normal conversation rather than a command, or it can interpret TV or background noise as a command and begin recording. These recordings can cause a backlog in the history file and decrease the accuracy in wake word identification. This study will analyze two Amazon Alexa voice recording datasets collected over a period of 45 days in an attempt to answer the research question: “which wake word yields the fewest inadvertent recordings.” The experimental network will consist of two Amazon Echo Dot’s, two Echo Show’s, and one Echo Tall, with all devices equipped with the Alexa voice software. Based on my findings, I will provide recommendations for wake word safe selection operation and possible amendments to future software updates as well as descriptive analysis of other voice recording characteristics such as of dialect recognition, tone of voice, and frequency of recordings.