Data Sources for Analysis of Voice in the Car 2019-01-0397
In vehicle voice recognition capabilities have become very prominent in the automotive industry with 55% of all new cars equipped with some form of voice recognition. These capabilities are intended to help customers engage various services like hands-free phone calling, navigation and others without taking their hands off the steering wheel and eyes off the road. However, recent studies indicate that many customers think that their systems don't recognize voice commands or interpret them correctly.
Information about how customers use voice recognition, under what circumstance, and the errors they encounter can be used to improve the system and direct the design of future voice recognition systems. Infotainment system logs, audio recordings of the voice interactions, their transcriptions and CAN bus data were identified to be rich sources of data to analyze voice recognition usage. This paper examines the importance of each data source and what questions related to customer usage of voice recognition they answer. Infotainment logs were found to be very useful in understanding how the system interpreted or responded to customer commands and at what confidence level. The logs can also help in identifying if the issued commands are of low confidence or have been cancelled by the user. The audio recordings of the voice interaction and their transcriptions provide information about what the customer actually said. The system’s interpretation of the command can be compared to the actual command to detect if it is correctly recognized by the system. Further, the transcriptions also help understand if customers issued commands that do not adhere to the grammar of the voice recognition system.
CAN bus data can be extremely useful in determining if voice recognition is used in the presence of other noise sources in the car such as HVAC blower, engine noise, etc. This makes it possible to determine if any voice recognition failures occur due to these noise sources in the car. Lastly, it was shown that all of these data sources can be tied together to understand the customers’ patterns of voice recognition usage which can be used to predict voice commands that may be issued in the future.
Ranjani Rangarajan, Scott Amman, Leah Busch