A METHOD FOR ASSESSING AUDIO FILE SIMILARITY USING AN AMT MODEL УДК 004.056.57

Abstract

The article discusses the development and experimental evaluation of a novelmethod for creating digital audio fingerprintsbased on Automatic Music Transcription (AMT)to enhance the efficiency of pirated contentdetection. The methodological foundationis an algorithm that converts an audio signalinto a symbolic representation (MIDI) usinga specially trained neural network model,"Google Onsets and Frames". This model,whose architecture includes two U-Net blocksand BiLSTM layers, was trained on the MAPSdataset of piano recordings. The Levenshteindistance was used to assess the similaritybetween the original and predicted MIDIrepresentations. Experimental results revealeda strong dependence of the model's accuracyon the nature of the audio data. The highestaccuracy (89.36%) was achieved on polyphonicpiano compositions that matched the trainingdata, while the results were significantly lower(35-56%) for monophonic and guitar tracks.To improve the method's reliability, a postprocessing algorithm filtering out false notesbased on a duration threshold was proposedand successfully tested. This procedureenabled achieving 100% similarity for mosttest piano recordings, proving the approach'seffectiveness. Thus, the study demonstrates the promise of using AMT for creating robustdigital fingerprints and outlines paths for furthermodel improvement, including expandingthe training dataset with data from variousinstruments.