Avtomatična transkripcija večglasnega petja

diplomsko delo

Uroš Paščinski (Author), Matija Marolt (Mentor)

Abstract

V tem delu se posvetimo avtomatični transkripciji večglasnega petja. Od celotne transkripcije se omejimo na detekcijo tonov. Pripravimo si testno množico večglasnih slovenskih ljudskih pesmi, ki jih pridobimo iz terenskih posnetkov, in zanje zgradimo ročne transkripcije. Nad testno množico poženemo splošni algoritem za polifonično detekcijo tonov. Razvijemo interaktivno vizualizacijo, ki nam ponazori delovanje algoritma in olajša iskanje napak. Rezultatov algoritma ne moremo primerjati, ker je testna množica nova. Algoritem poskusimo izboljšati. Utežitveno funkcijo magnitudnega spektra nadomestimo z linearno, kar prinese slabše rezultate od prvotnih. Poskusimo tudi z dvakratnim beljenjem magnitudnega spektra, ki se obnese nekoliko bolje, a še vedno ne zadovoljivo. Z mehkejšim pristopom k vrednotenju opazimo, da je lahko vzrok težav slabe detekcije tudi problematična testna množica, saj ima precej intonančnih težav.

Keywords

avtomatična transkripcija;večglasno petje;slovenske ljudske pesmi;interaktivna vizualizacija;terenski posnetki;računalništvo;računalništvo in informatika;računalništvo in matematika;interdisciplinarni študij;univerzitetni študij;diplomske naloge;

Data

Language:	Slovenian
Year of publishing:	2015
Typology:	2.11 - Undergraduate Thesis
Organization:	UL FRI - Faculty of Computer and Information Science
Publisher:	[U. Paščinski]
UDC:	004:78.087.6(043.2)
COBISS:	1536278211
Views:	1959
Downloads:	209
Average score:	0 (0 votes)
Metadata:

Other data

Secondary language:	English
Secondary title:	Automatic transcription of polyphonic singing
Secondary abstract:	In this work we focus on automatic transcription of polyphonic singing. In particular we do the multiple fundamental frequency (F0) estimation. From the terrain recordings a test set of Slovenian folk songs with polyphonic singing is extracted and manually transcribed. On the test set we try the general algorithm for multiple F0 detection. An interactive visualization of the main parts of the algorithm is made to analyse how it works and try to detect possible issues. As the data set is new we cannot compare the results. Steps are made towards improvements of the algorithm. The magnitude spectrum weighting function is replaced with a simple linear function but results in the degradation of the performance. Then we try to use double spectral whitening of the magnitude spectrum which turns out more promising, but still not satisfactory. A softer evaluation criteria shows that errors in performance might be due to the problematic test set, which has lots of intonation errors.
Secondary keywords:	automatic transcription;polyphonic singing;Slovenian folk songs;interactive visualization;terrain recordings;computer science;computer and information science;computer science and mathematics;interdisciplinary studies;diploma;
File type:	application/pdf
Type (COBISS):	Undergraduate thesis
Thesis comment:	Univ. v Ljubljani, Fak. za računalništvo in informatiko
Pages:	101 str.
ID:	8752166