Binaural conversion of monaural music recordings using head-related impulse responses

This Letter investigates the interplay between source-separation quality and head-related impulse response (HRIR)-based binaural rendering for monaural music recordings to provide design guidelines for monaural-to-binaural processing. Instrument signals are separated from monaural mixtures using nonnegative matrix factorization and, then, listener-specific HRIRs are convolved with separated sources to synthesize binaural audio in which each instrument is assigned a distinct direction. Subjective listening tests indicate that when source-separation accuracy is sufficiently high, the HRIR-based conversion improves localization accuracy while maintaining comparable sound quality and localization stability. Results also reveal how residual components from imperfect source separation degrade binaural rendering.