Voice interaction is friendly and convenient for users. Smart devices such as Amazon Echo allow users to interact with them by voice commands and become increasingly popular in our daily life. In recent years, research works focus on using the microphone array built in smart devices to localize the user's position, which adds additional context information to voice commands. In contrast, few works explore the user's head orientation, which also contains useful context information. For example, when a user says, "turn on the light", the head orientation could infer which light the user is referring to. Existing model-based works require a large number of microphone arrays to form an array network, while machine learning-based approaches need laborious data collection and training workload. The high deployment/usage cost of these methods is unfriendly to users. In this paper, we propose HOE, a model-based system that enables Head Orientation Estimation for smart devices with only two microphone arrays, which requires a lower training overhead than previous approaches. HOE first estimates the user's head orientation candidates by measuring the voice energy radiation pattern. Then, the voice frequency radiation pattern is leveraged to obtain the final result. Real-world experiments are conducted, and the results show that HOE can achieve a median estimation error of 23 degrees. To the best of our knowledge, HOE is the first model-based attempt to estimate the head orientation by only two microphone arrays without the arduous data training overhead.
Sound Localization for Ad-Hoc Microphone Arrays