Transcription of piano music with deep learning

Transkripcija glasbe je zahteven postopek simboličnega zapisa glasbenega posnetka. Cilj tega diplomskega dela je bila preučitev transkripcije klavirske glasbe z metodami globokega učenja, za kar so bili implementirani in preizkušeni trije modeli globokih nevronskih mrež: večnivojski perceptron, konvolucijska nevronska mreža in globoka verjetnostna mreža. Z modelom globoke verjetnostne mreže je bilo preizkušeno nenadzorovano predučenje, katerega namen je izluščenje glasbenih značilnosti iz zvočnega signala. Učenje modelov in preverjanje končne uspešnosti transkripcije je bilo izvedeno na zbirki za transkripcijo klavirske glasbe MAPS. Izvedena je bila tudi primerjava predpriprave podatkov s transformacijama hitre Fourierove transformacije in transformacije s konstantnim Q. Končni rezultati so pokazali, da je globoko učenje s pravim učnim načrtom lahko močno orodje za transkripcijo glasbe.Transcription of music is a complex process of transcribing an audio recording into a symbolic notation. The goal of this thesis was to examine transcription of piano music with deep learning, for which three models of deep neural networks were implemented: multilayer perceptron, convolutional neural network and deep belief network. Through the use of deep belief network, unsupervised pretraining for automatic extraction of musical features from audio signals was also tested. Learning of these models and evaluation of transcription was performed with MAPS database for piano music transcription. A comparison between Fast Fourier Transform and Constant Q Transform for data pre-processing was also carried out. Final results show that deep learning with an appropriate learning schedule is potentially a powerful tool for automatic transcription of music