A voice interaction architecture that compiles multiple audio signals captured at different locations within an environment, determines a time offset between a primary audio signal and other captured audio signals and identifies differences between the primary signal and the other signal(s). Thereafter, the architecture may provide the primary audio signal, an indication of the determined time offset(s) and the identified differences to remote computing resources for further processing. For instance, the architecture may send this information to a network-accessible distributed computing platform that performs beamforming and/or automatic speech recognition (ASR) on the received audio. The distributed computing platform may in turn determine a response to provide based upon the beamforming and/or ASR.