Digital Signal Processing Reference
In-Depth Information
The onsets are likewise represented by the local maxima of
o
o
(
)
n
. With a standard
(
)
peak search, the final onset function
o
n
is:
1for
o
o
(
n
−
1
)
≤
o
o
(
n
)
≥
o
o
(
n
+
1
)
o
(
n
)
=
(11.13)
0
otherwise
11.2.3 Performance
In the literature, an onset is correctly located when detected
±
50 ms [
52
,
56
]or
±
25 ms [
62
] around the annotated gold standard onset position. Results with a fixed
threshold scaling factor of
λ
=
50 are given per set for both of these two tolerance
criteria. Note that humans are believed to perceive two onsets as one if they are no
more than 30 ms apart [
63
]. On the Bello set, this would in theory leave 4294 from
the 5474 onsets that can be distinguished by humans.
Evaluation on
BRD
o
and the Bello set bases on eight-fold SCV where six folds
are used for training, one for development, and one for testing. Owing to the random
initialisation of BLSTM RNN, the eight-fold cross validation is repeated ten times
Fig. 11.4
To p
log Mel-spectrogram with ground truth onsets (
vertical dashed lines
).
Bottom
net-
work output with detected onsets (marked by
dots
), ground truth onsets (
dotted vertical lines
), and
threshold
θ
(
horizontal dashed line
). Shown is a 4 s excerpt from 'Basement Jaxx—Rendez-Vu' [
23
]
