A web tool for comparison of music transcription systems

V diplomski nalogi razvijemo spletno orodje, ki omogoča primerjavo algoritmov za samodejno transkripcijo glasbe. Preko grafičnega vmesnika lahko primerjamo časovno zahtevnost izvajanja ter grafično predstavitev in primerjavo rezultatov transkripcij. Nad izbranimi algoritmi omogoča računanje in primerjavo glavnih evalvacijskih mer. Orodje razvijemo na osnovi štirih algoritmov. Na podlagi te vhodne množice aplikacijo posplošimo z možnostjo dodajanja novih. S tem povečamo uporabnost orodja, saj uporabniku omogočimo primerjavo rezultatov lastnih algoritmov s tistimi, ki so v sistem že dodani. Med razvojem aplikacije algoritme izvajamo na zbirki studijskih klavirskih posnetkov MAPS. Ker tovrstne zbirke poleg vsakega zvočnega posnetka vsebujejo še referenčno anotacijo, to izkoristimo ter omogočimo prikaz natančnosti transkripcije posameznega algoritma. Aplikacija omogoča tudi nalaganje novih zvočnih zbirk. Proces samodejne transkripcije glasbe je časovno zahtevnejša operacija, zato dodamo uporabniku možnost, da ob zaključku procesa prejme obvestilo v elektronski poštni predal. Poleg navedenega nam orodje omogoča še grafično urejanje in izvoz rezultata transkripcije v zapisu MIDI. Z omenjenim naborom funkcij na področje raziskovanja samodejne transkripcije glasbe doprinesemo z uporabnim in enostavnim orodjem.This thesis presents the development of a web application that enables the user to compare algorithms for automatic music transcription. The user interface provides visualization of time difficulty, graphic representation and comparison of the algorithm’s transcription results. Furthermore, it enables calculation and comparison of main evaluation metrics. The system was developed by using four different transcription algorithms. Based on this input set, the application was generalized by giving the user the ability to add his or her own solutions. This functionality significantly increased usability by providing a comparison of user algorithm results with those already available in the system. For algorithm execution, an annotated sound database, MAPS was used. These sorts of data sets include ground-truth transcriptions for each audio file, therefore adding an extra option to evaluate the algorithm’s result accuracy. The application provides a file upload function as well, in order to extend an annotated sound library. Automatic music transcription is generally a time-consuming process, hence an option that notifies the user of its termination had been added. In addition to all functions stated, this tool provides editing and exporting transcription results in MIDI format. With this set of options a simple and useful tool had been introduced to the automatic music transcription community