Estimating the Direction of Arrival of a Spoken Wake Word Using a Single Sensor on an Elastic Panel

The vibrations induced in an elastic panel from an incident acoustic pressure wave are a function of the resonant mode structure of the panel and the angle of incidence of the acoustic wave. In this paper it has been demonstrated how measurement of the panel’s modal response with a single structural vibration sensor may be employed to infer the direction of arrival (DOA) of the incident sound. The method is dependent on the frequency content of the acoustic wave, as modes that provide important spatial information about the source may not be excited if the acoustic signal lies outside their resonant bandwidths. This work explores techniques for extending this single-sensor approach to DOA estimation for speech signals, which represent a realistic use case for applications such as smart audio devices. Feature sets including Mel spectrograms, Mel-frequency cepstral coefficients (MFCCs), and linear spectrograms, were used to train convolutional and feedforward neural networks to estimate the DOA of a wake word recorded by a single structural vibration sensor affixed to a panel. The experiments were carried out in semi-anechoic conditions and are thus presented as proof of concept. Additionally, the models presented are compact enough to be deployed on embedded/edge hardware commonly used in smart audio devices. The trained models estimated the DOA of the wake word utterance to within ±5° with an average reliability of 83.1% when using MFCCs as features. This average reliability improved to 92.23%, with a maximum reported reliability of 99.9%, when using Mel and magnitude spectrograms and an additional hardware-specific feature set, suggesting that single-sensor DOA estimates for speech signals may be improved by using more spectrally complete feature sets.