Embedded Files
At Microsoft, I was responsible for voice AI development and infotainment optimization.
Video Demo showcases the command "close the door".
2 things that I am checking:
Whether the door successfully closed (passed)
Whether the speech feedback from the system is the correct answer (failed)
This command is considered a basic command (highest product development priority) and needs to be ~100% accurate. In-person tests were done in addition to automated tests.
Azure AI CLU
For the in-vehicle voice AI, the system is connected to the internet for navigation, playing music, finding service centers/chargers, etc.
On the right is a typical test case of voice recognition on Azure AI's Conversational Language Understanding (CLU). I re-architected the labeling strategy for the online language model of voice AI, specifically fixing the contextual problem of the model mistaking everything to be a song name (i.e., "window" should be recognized as an entity of the car's window, but the model can judge "window" to be a song called window).
Product Requirement Documents
I composed product requirement documentation for this manufacturer's updated center infotainment and instrument cluster. The PRDs consisted of product development priorities, feature details, and workflow. Most of the features were designed based on historical user feedback, interaction data, and internal tests. Each newly designed feature was integrated into the voice AI in the next version update.
Voice AI Testing
In addition to running automated tests, I performed automated testing on Android Studio. I always first run a standard test case a few times, then run randomly generated test cases. If the percentage accurately met the criteria, a new version would be pushed to Git. After pushing a new version, Trial tests would be performed on the emulator.
NOTE: This is a company project and limited information can be shared. No proprietary information is displayed.
Page updated
Google Sites
Report abuse