Microphone Array Coding Preserving Spatial Information for Cloud-based Multichannel Speech Recognition

An efficient method of coding multichannel signals from a microphone array is presented. This is advantageous for cloud-based audio processing, such as Direction-of-Arrival (DOA) and Automatic Speech Recognition (ASR). The method operates by encoding separately the signal information - using a reference signal - and the spatial information - using Relative Transfer Functions (RTFs). Results for ASR and DOA performance are presented for the proposed codec in comparison to a baseline multichannel implementation of the Opus codec. Both stationary and time-varying acoustic scenarios have been included in the tests. The proposed RTF-based codec is shown in our experiments to preserve spatial information in the array signals whereas the baseline codec does not. The proposed codec is also shown to outperform the baseline on the ASR task at low bit rates in the region of 6 kbits per second per channel.