Digital Signal Processing Reference
In-Depth Information
-
Fig. 6.11
Incremental data-flow in openSMILE's ring-buffer memories on LLD level. The (
light
)
arrows
pointing in between the columns depict the current write pointer [
94
]
Figure
6.12
next shows this incremental processing of higher order features such
as functionals to project the time series to single feature values. Shown are two exem-
plary functionals, namely 'max' and 'min'. These are calculated over two overlapping
frames from the pitch LLD. Then, they are saved to the level 'func'. The buffers-size
is matched to the block-size of the reader or writer. In the pitch functionals example
the read block-size of the functionals component thus would be two because two pitch
frames are read at once. openSMILE supports multi-threading for fast computation.
For utmost parallelisation on multi-core computers, each component can be run in a
separate thread. Individual components can further be freely instantiated, configured,
and connected to the
Data Memory
via a central configuration file. Further, on-line
audio recording and live feature extraction is possible.
6.5.2 Available Feature Extractors
openSMILE provides a number of LLDs (cf. Table
6.1
) for automatic extraction
and the application of several filters, functionals, and transformations to these. Mel-
spectra, MFCCs, and PLPs can be computed exactly in full compliance with the
popular Hidden Markov Toolkit (HTK) [
20
], fostering compatibility and compara-
bility. PLP computation can be carried out as in original works [
22
] or in modification
(eg leaving out processing steps).