| Meno: | Jozef
|
|---|
| Priezvisko: | Budáč
|
|---|
| Názov: | Polyphonic music transcription via deep neural networks
|
|---|
| Vedúci: | prof. RNDr. Martin ©koviera PhD.
|
|---|
| Rok: | 2020
|
|---|
| Kµúčové slová: | automatic music transcription, deep learning, MAPS, MIDI
|
|---|
| Abstrakt: | We explore the field of automatic polyphonic music transcription by using a deep convolutional neural network which is trained to predict pitches of both classical piano music and Slovak folklore music. As a source of inspiration, we take an article published by Sigtia et al., and we replicate acoustic models used in their work. We improve our replicated Sigtia ConvNet acoustic model by using popular contemporary machine learning concepts in order to get closer to their results. Our approach results in over a 9% relative improvement in our experiments on the MAPS dataset. Furthermore, we create a novel Multi-instrument midi aligned sounds dataset focused on traditional Slovak music to improve the performance of predictions on Slovak folklore multi-instrumental records.
|
|---|