MICHAEL BLAß AND ROLF BADER
CONTENT-BASED MUSIC RETRIEVAL SYSTEM FOR ETHNOMUSICOLOGICAL SOUND ARCHIVES
Preservation All photos by Christian Koehn,
[email protected]
Exploration All photos by Christian Koehn,
[email protected]
1. EXAMPLE APPROACHES TO
ETHNOGRAPHIC SOUND ARCHIVES
1. The classical approach: •
Archiv für die Musik Afrikas (AMA) in Mainz, Germany •
catalogue based
•
about 10,000 sound carriers of modern African music
•
digitization in progress
•
+ exploration
•
- not feasible for big archives
We argue, that even though digital musicology already claims to be bridging a gap between “fieldwork ethnomusicologists” and “big data, experts”, that a combination, the best of both worlds may be a profitable way to go. You cannot explore an music archive without using your ears.
1. EXAMPLE APPROACHES TO
ETHNOGRAPHIC SOUND ARCHIVES
2. The digital musicology approach: •
Digital Music Lab (DML), London, U.K. •
purely feature based
•
analyze and compare songs or even
whole archives on a variety of sound
features
•
read only, no sound
1. EXAMPLE APPROACHES TO
ETHNOGRAPHIC SOUND ARCHIVES ➡ We need the best of both worlds
1. Catalogue / text based search and retrieval engine •
online access
•
access to audio
2. Content-based, automatic archive organization •
sound features
•
music similarity
2. Ethnographic sound recordings Archive (ESRA) University of hamburg
2. Ethnographic sound recordings Archive (ESRA) University of hamburg
•
The Wilhelm Heinz Collection of African Music •
350 piece
•
mostly gramophone records
•
1916 – 1948
•
38 regions
•
49 ethnics
Yildiz Quantasi Thessaloniki, ca. 1920
2. Ethnographic sound recordings Archive (ESRA) University of hamburg
•
The Cairo Congress of Arab Music •
103 piece
•
March–April 1932
•
Algeria, Tunisia, Iraq, Turkey, Syria
Title: ya Muqabil Performer: al-Gawq al-Gaza'ir 1932
2. Ethnographic sound recordings Archive (ESRA) University of hamburg
•
Rolf Bader Field Recordings •
118 piece
•
2010 until now
•
Myanmar, Sri Lanka, Cambodia
3. CHALLENGES IN ETHNOGRAPHIC SOUND ARCHIVES
•
Meta data
3. CHALLENGES IN ETHNOGRAPHIC SOUND ARCHIVES
3. CHALLENGES IN ETHNOGRAPHIC SOUND ARCHIVES
3. CHALLENGES IN ETHNOGRAPHIC SOUND ARCHIVES PLENT Y OF CL ASSES •
Meta data
•
Plenty of classes
•
Unbalanced data
3. CHALLENGES IN ETHNOGRAPHIC SOUND ARCHIVES UNBAL ANCED DATA
3. CHALLENGES IN ETHNOGRAPHIC SOUND ARCHIVES UNBAL ANCED DATA
3. CHALLENGES IN ETHNOGRAPHIC SOUND ARCHIVES
•
Meta data
•
Plenty of classes
•
Unbalanced data
•
heterogenous data
3. CHALLENGES IN ETHNOGRAPHIC SOUND ARCHIVES
•
Meta data
•
Plenty of classes
•
Unbalanced data
•
heterogenous data
3. CHALLENGES IN ETHNOGRAPHIC SOUND ARCHIVES
•
Meta data
•
Plenty of classes
•
Unbalanced data
•
heterogenous data
•
noise
4. COMSAR COMPUTATIONAL MUSIC AND SOUND ARC HIVING PROJECT
•
Provide rich meta data structure
•
Visualization by music similarity
COMSAR COMPUTATIONAL MUSIC AND SOUND ARC HIVING PROJECT
•
extract low-level timbral feature
•
aggregate to song-level rhythm feature
•
visualization mutual similarity
COMSAR RHY THM FEATURE
(Blaß, 2013a; Blaß 2013b)
COMSAR RHY THM FEATURE
•
Identical in IOI
•
different in timbre
•
Can be generalized to polyphonic timbre
(Blaß, 2013a; Blaß 2013b)
TIME DOMAIN SIGNAL AND SPECTROGRAM Section: 0 Hz – 10 kHz
STFT parameters: n_fft = 1024, hop = 512, Hamming window
TIMBRE FEATURE SPECTRAL CENTROID OR MFCC? •
Mel frequency cepstral coefficients (MFCC) •
•
Primary feature for timbre related computations
Spectral Centroid •
Multidimensional scaling (MDS) of dissimilarity ratings
for example, Grey (1977); Wessel (1979); Inverson & Krumhansel (1993)
•
MDS of physical signal parameters
for example, Lakatos (2000)
•
Adjective rating studies
for example, von Bismark (1977)
FEATURE EXTRACTION TIMBRE SPACE FROM GREY (197 7)
FEATURE EXTRACTION TIMBRE SPACE FROM GREY (197 7)
FEATURE EXTRACTION SPECTRAL CENTROID AS MODEL FOR TIMBRE •
Previous studies confirmed Spectral Centroid as main perceptual dimension.
•
Spectral Centroid correlates well with perception of brightness (Schubert et al, 2004).
•
Mel-frequecy cepstral coefficients do not seem to correlate with any perceptual dimension (Alluri & Toiviainen, 2009; Siedenburg et al., 2017).
SYSTEM OVERVIEW 🎶 Source
Preprocessing
Feature extraction
Clustering/ Visualization
➡
✄ SC
ONSET DETECTION BASED ON SPECTRAL FLUX
SF (n) =
PK
1 k=0
H (|X(n, k)| |X(n PK 1 k=0 |X(n, k)|
1, k)|)
x + |x| H(x) = 2
X = Spectrum n = Number STFT window k = Number frequency bin
SYSTEM OVERVIEW 🎶 Source
Preprocessing
Feature extraction
Clustering/ Visualization
➡
✄ SC
TIMBRE FEATURE SPECTRAL CENTROID
SC(n) =
K X1
f (k) p(n, k)
k=0
X(n, k)
p(n, k) = PK
1 k=0
X(n, k)
X = Spectrum n = Number STFT window k = Number frequency bin
SYSTEM OVERVIEW 🎶 Source
Preprocessing
Feature extraction
Clustering/ Visualization
➡
✄ SC
CONCLUSION
•
Ethnographic sound archives have a strong potential that goes far beyond mere conservation.
•
Currently there is a lot of effort to utilize this potential.
•
Classical methods support the exploration approach but are not adequate for big archives
•
Big data methods are abstract
CONCLUSION
•
•
We propose a data drive (content-based) system that can support ethnomusicological research by … •
ordering existing music archives
•
finding new hypotheses
The system consists of 1. Hidden Markov Model (HMM)-based feature extraction, which models rhythm in terms of occurring polyphonic timbres. 2. Self-Organizing Map (SOM) projection, which orders the HMMs to rhythmically similar clusters.
CONCLUSION •
•
Evaluation is very difficult: •
Archives would have to be annotated in order the measure
onset detection performance.
•
There is not target data about rhythm
•
SOM is a tool for explorative data analysis.
Evaluation by expert users.
VISIT AND TRY ESRA
IT’S FREE
http://esra.fbkultur.uni-hamburg.de
THANK YOU for your attention.
“Content-based music retrieval system for ethnomusicological sound archives” Michael Blaß
[email protected]
Rolf Bader
[email protected]
Visit ESRA http://esra.fbkultur.uni-hamburg.de
We appreciate your feedback, especially regarding ESRA. Michael Blaß, M.A.
Prof. Dr. Rolf Bader
[email protected] http://www.uni-hamburg.de/ifsm
[email protected] http://www.uni-hamburg.de/ifsm
Institute for Systematic Musicology University of Hamburg Neue Rabenstraße 13 20354 Hamburg Phone: +49 40 42838-5786