Using a Physiological Ear Model for Automatic Melody Transcription and Sound Source Recognition

Recent trends in musical audio signal analysis increasingly promote use of perceptual motivated modifications of conventional signal processing algorithms. A consequent further step consists of the inclusion of the knowledge of the structure of mammalian auditory periphery. The presented work uses physiological models in order to mimic active functionality of the inner ear including the transduction from mechanical vibrations into neural impulses. The main part of the paper describes automatic transcription of melodies from real world musical inputs. Bottom-up extraction and segmentation of pitch trajectories based on the outputs of the used models, i.e.~concentration of transmitter substance inside the inner hair cell clefts, are demonstrated. As an example for the wide range of possible further applications a sound source recognition approach using woodwind instruments is proposed. Results indicate that the algorithm performs excellent compared to traditional methods.