Microphone array processing for distant speech recognition: Spherical arrays

摘要

Distant speech recognition (DSR) holds out the promise of the most natural human computer interface because it enables man-machine interactions through speech, without the necessity of donning intrusive body- or head-mounted microphones. With the advent of the Microsoft Kinect, the application of non-uniform linear arrays to the DSR problem has become commonplace. Performance analysis of such arrays is well-represented in the literature. Recently, spherical arrays have become the subject of intense research interest in the acoustic array processing community. Such arrays have heretofore been analyzed solely with theoretical metrics under idealized conditions. In this work, we analyze such arrays under realistic conditions. Moreover, we compare a linear array with 64-channel arrays and a total length of 126 cm to a spherical array with 32 channels and a radius of 4.2 cm; we found that these provided word error rates of 9.3% and 10.2%, respectively, on a DSR task. For a speaker positioned at an oblique angle with respect to the linear array, we recorded error rates of 12.8% and 9.7%, respectively, for the linear and spherical arrays. The compact size and outstanding performance of the spherical array recommends itself well to space-limited and mobile applications such as homegaming consoles and humanoid robots.