Evaluation of the

peccarytibburΤεχνίτη Νοημοσύνη και Ρομποτική

25 Νοε 2013 (πριν από 3 χρόνια και 11 μήνες)

105 εμφανίσεις

Evaluation of the
Audio Beat Tracking
System
BeatRoot

By Simon Dixon (JNMR 2007)


Presentation by
Yading

Song

Centre for Digital Music

y
ading.song@eecs.qmul.ac.uk


QMUL

ELE021

Music

&

Speech

Processing

27

February

2012


Identifying and synchronizing with the basic rhythmic pulse of
a piece of
music


An interactive beat tracking and metrical annotation system[1]


It uses a multiple agent architecture with different hypotheses


Rate


Placement of musical
beats


Accurate tracking


Quick recovery from errors


Graceful degradation

BeatRoot



Tempo induction


F
ind the rate of beat


Beat tracking


Synchronize a quasi
-
regular pulse sequence with music


Steps

Architecture of
BeatRoot

System


Onset Detection



Tempo Induction



Beat Tracking





Detection function


Spectral flux (used by Dixon)


Weighted phase deviation


Complex domain detection function


Spectral Flux


The square of the difference between the normalized
magnitude of successive frames


How quickly the power spectrum of the a signal is
changing


Peak
-
picking algorithm is used to find the local maxima


Onset detection function




Onset Detection

Spectral Flux

Example of spectral flux “
vivaldi.wav
”, implemented in
MIRtoolbox



It calculates onsets times to compute clusters of
inter
-
onset intervals (IOIs)


IOI = the time interval between any pair of onsets


Use clustering algorithm to find groups of similar IOIs


Represents various musical units (e.g. half notes)


Tempo Induction



1. Clustering


Various of IOIs


Greedy algorithms




2. Combining


Along with the No. of IOIs


To weight the clusters


A ranked list of tempo hypotheses is produced


Pass it to beat tracking sub
-
system

Two steps



It uses a multiple agent architecture to find sequence of
events


Match various tempo hypotheses


Rate each sequence


Determine the most likely one


The music is processed sequentially from beginning to
end


At any point the agents


Represent various hypotheses about the rate and timing of beat


Make prediction of next beats based on current states




Beat Tracking


Each agent at the beginning


Is initialized with a tempo hypothesis


An onset time which is taken from the first few onsets, which
defines the agent’s first beat time


Make prediction with given tempo and first beat time with a
tolerance window





Onsets


In inner window


taken as actual beat time, stored and updated


In outer window


taken as possible beat times or not


Beat Tracking


Beat Tracking

Solid circle: predicted beat times which correspond to onset

Hollow circle: predicted beat times which don’t correspond to onset


Each agent is equipped with an evaluation function which
rates how well the predicted and actual beat correspond


The agent with the highest score outputs sequence of beats as
the solution to the beat tracking problem


Beat Tracking


User Interface

User Interface


Tempo Induction is correct in the most case


Estimation of beat times
are robust [2]


Evaluation

[1]
S. Dixon, "Evaluation of audio beat tracking system
beatroot
,"
Journal

of
New Music Research
, vol. 36, no. 1, pp.
39
-
51, 2007.

[2]
MIREX, Music Information Retrieval Evaluation
eXchange


Reference


Yading Song

Centre
for Digital Music

yading.song@eecs.qmul.ac.uk


Comments?