Learning music time series with a parallel hybrid forecasting model calibrated with Reduced VNS

This paper focus on the learning of music time series. In this context, from compressed digital audio files, we sought to verify how a song can be learned, both in terms of amplitude and frequency. Given the enormous amount of data contained in those time series, the use of classical learning methods becomes limited. Typical compressed acquisitions in MP3 files usually contains 44100 samples per second. In this context, the use of metaheuristic algorithms for this learning task, in a big-data environment, is justified, and the use of deep learning techniques sounds necessary. In this paper, a Hybrid Forecasting Model, calibrated with the Reduced Variable Neighborhood Search, with parallel processing using Graphical Processing Units, is used as a deep learning tool. Case studies composed of simple musical compositions are used for verifying the potential of the method for such application. We suggest that the techniques investigated here can also be used for the learning, classification and computational music composition.