Digital Signal Processing Reference
In-Depth Information
s(k)
Preprocessing
Mel-bands, envelopes, down-sampling
s'(k)
Feature Extraction
comb-filter bank
x
Metre Classification
SVM
Metre (duple/ triple)
Ballroom Dancestyle Classification
SVM
Dancestyle (9 classes)
Tempo Detection
exploitation of tempo / tatumstatistics
Tempo Detection
Beat leveltempo (BPM)
Fig. 11.11
Flow in the described data-driven tempo detection basing on metre and ballroom dance
style recognition [
6
]
Figure
11.11
shows the overall processing flow for metre, ballroom dance style,
and quarter-note tempo determination: First, a SVM-model for metre classification
is built using the feature sub-set
F
metre
to assign a metre
M
(duple or triple). Then,
the metre
M
is used as a feature in the set
F
dance
(cf. Sect.
11.3.2.3
) for ballroom
dance style classification. Finally, determined metre and ballroom dance style are
used to assess quarter-note tempo robustly.
From the training data the means
2
q
μ
q
/
T
and variances
σ
T
of the annotated
/
quarter-note tempi and tatum tempi
θ
T
are calculated per ballroom dance styles. As
no annotation for tatum tempo is usually available, the tempo estimated automatically
as in the first step (cf. Sect.
11.3.2.1
) serves as substitute. Higher WA could be reached
given manual annotation also for this tempo.
Then, the tempo of unknown test instances is determined: With the two tatum
candidates
θ
T
2
as extracted in the first step in Sect.
11.3.2.4
, the final tatum
is decided upon based on the statistics from the training data. The confidence
C
T
1
/
2
θ
T
1
and