E-Locus - Institutional Repository of the University of Crete

Home Collections Type of Work Post-graduate theses

Post-graduate theses

Current Record: 4 of 6633

[Add to Basket]

Identifier

000464520

Title

Application of similarity learning and few-shot learning techniques for percussive sound recognition

Alternative Title

Εφαρμογή τεχνικών μάθησης ομοιότητας και μάθησης με λίγα παραδείγματα για αναγνώριση κρουστικών ήχων

Author

Συμιακάκης, Ανδρέας Κ.

Thesis advisor

Τσακαλίδης, Παναγιώτης

Reviewer

Στυλιανού, Γιάννης
Στεφανάκης, Νικόλαος

Abstract

The study of hyper-instruments or augmented musical instruments is a very fascinating topic in music technology. It involves the augmentation of the capabilities of traditional musical instruments - as a source of music creation and performance - by integrating technologies that, until recently, were mostly known in the context of electronic musical instruments. Advances in computing and machine learning (ML) have made it possible to develop hyper-instruments that do not require invasive, irreversible modifications on the instrument, or additional equipment such as sensors or controls that musicians are unfamiliar with. However, training machine learning models for such a purpose requires a large amount of unusual data that can be time-consuming to acquire and that is not usually publicly available. This Thesis studies the application of machine learning techniques in the context of a guitar augmentation system, assuming a user who interacts with the system through percussive gestures produced on the body of the guitar. It studies metric learning and few-shot learning techniques and discusses the value of these techniques on the basis of numerous experiments with the aim of (i) finding the optimal approach for training a siamese encoder using a large set of labelled sound examples from different acoustic or classic guitars and (ii) finding the best way to exploit the pre-trained siamese encoder for recognising a set of unknown samples from a new guitar, given a limited only support set from that new guitar. Experimental results show that the proposed combination of ML techniques provides a viable approach in the context of few-shot learning, enabling a significant reduction in the preparation time and effort required from the final user.

Language

English

Issue date

2024-07-26

Collection

School/Department--School of Sciences and Engineering--Department of Computer Science--Post-graduate theses

Type of Work--Post-graduate theses

Permanent Link

https://elocus.lib.uoc.gr//dlib/0/1/1/metadata-dlib-1715239001-990128-4184.tkl

Views

Digital Documents
	Download document View document Views : 1