X-net: A Joint Scale Down and Scale Up Method for Voice Call

This paper proposes X-net, a jointly learned scale-down and scale-up architecture for data pre- and post-processing in voice calls, as a means to bandwidth extension over band-limited channels. Scale-down and scale-up are deployed separately on transmitter and receiver to perform down- and upsampling. Separate supervisions are used on the submodules so that X-net can work properly even if one submodule is missing. A two-stage training method is used to learn X-net for improved perceptual quality. Results show that jointly learned X-net achieves promising improvement over blind audio super-resolution by both objective and subjective metrics, even in a lightweight implementation with only 1k parameters.