Digital Signal Processing Reference
In-Depth Information
Intonation
((multiple) pitch, …)
Intensity
(energy, Teager, …)
Linear Predicition
(LPCC, PLP, ...)
CepstralCoefficients
(MFCC, HFCC, …)
Formants
(amplitude, position, …)
Spectrum
(PCP, CHROMA, ...)
TF-Transformation
(wavelets, Gabor, …)
Harmonicity
(HNR, NHR, ...)
Pertubation
(jitter, shimmer, …)
Linguistics
(phonemes, chords, …)
Non-Linguistics
(laughter, sighs, …)
Extremes
(min, max, range, …)
Means
(arithmetic, absolute, …)
Percentiles
(quartiles, ranges, …)
Higher Moments
(std. dev., kurtosis, …)
Peaks
(number, distances, …)
Segments
(number, duration, …)
Regression
(coefficients, error, …)
Spectral
(DCT coefficients, …)
Temporal
(durations, positions, …)
VectorSpace Modelling
(bag-of-words, …)
Look-Up
(wordlists, concepts, …)
Derving
(rawLLD,
deltas, regression
coefficients,
correlation
coefficients,
…)
Acoustics
(numeric)
Deriving
(rawfunctionals,
hierarchical,
cross-LLD,
cross-chunking,
contextual,
…)
Deriving
(rawLLD,
stemmed, POS-,
semantic
tagging, …)
Linguistics
(symbolic)
Disfluencies
(pauses, …)
Statistical
(salience, infogain, …)
Low-Level Descriptors
Functionals
Fig. 6.13
Overview on the principle of audio feature brute forcing in several hierarchical lay-
ers.These are generally divided into LLD and the subsequent (optional) Functional level. Shown
are further acoustic and linguistic features
the value or contribution of features or feature groups. Examples of filter functions
are statistic and information theoretic measures such as CC or IGR.
Given the size of the data set and the feature space, a search algorithm and simple
evaluation or ranking functions may additionally become mandatory, as exhaustive
search of all possible feature combinations can become computationally prohibitive.
A simple, yet highly efficient search method is 'conservative hill climbing', i.e.,
sequentially deciding for the best feature at the time starting from one and adding the
'next best', each. As this obviously is prone to nesting effects, one usually adds a back
stepping option whether 'another previous candidate' would have better suited. This
is known as 'floating search', and with the described forward addition as Sequential
Forward Floating Search. A backward search starting from the full feature set as
well as bi-directional searches are alternatives depending on the ratio of the feature
inventory and the target space size. As a result of a typical search, one obtains a mixed
view as for the brute-forced features, which is usually hard to interpret: Features in
the 'optimal' set, are usually a mixture of all groups. Yet, it is not clear whether these
are the best due to the suboptimal nature inherent in any search function and the
fact that it de-correlates the space rather than ranks. By that, the value of a feature
is unclear, as is whether a picked feature does not have a counter-part of similar
characteristics that was not picked, as only one of a sort is needed. An alternative is
a systematic 'scan' by feature groups, for examples per LLD type and per functional
type.
