![](https://csdnimg.cn/release/download_crawler_static/15951637/bg4.jpg)
T. Lu et al.
1 3
classification. Ghoraani and Krishnan [29] construct the
time–frequency matrix (TFM) of audio signals and apply
the non-negative matrix decomposition to decompose TFM
into its significant components. As a result, they propose
seven new features from spectral and temporal structures
of the decomposed vectors in a way that they success-
fully represent joint time–frequency (TF) structure of the
audio signal, and combine them with the MFCCs features.
Unfortunately, adding more features is not always helpful
as illustrated in [19], where the use of all features does not
always produce good performance for the audio classifica-
tion problem. It is probably due to the fact that as the fea-
ture dimension increases, data points become sparser and
potentially irrelevant features could negatively impact the
classification result.
It in turns leads to another strategy of choosing an opti-
mal subset of features from a larger set of possible fea-
tures to yield more accurate and the most effective subset.
Umapathy et al. [30] propose a time–frequency approach
for audio classification, which is considered the best way
to analyze audio signals non-stationary in nature. In their
method, audio signals are decomposed using an adaptive
TF decomposition algorithm to generate a set of 42 fea-
tures over three frequency bands within the auditory range.
These features are analyzed using linear discriminant func-
tions and classified into six groups. In their succeeding
work [31], they further propose an audio feature extrac-
tion and a multigroup classification scheme that focuses
on identifying discriminatory time–frequency subspaces
using the LDB technique. Chu et al. [32] propose to use
the matching pursuit algorithm to select effective time–fre-
quency features as the supplementation to MFCC for audio
environment characterization and general acoustic scene
types are represented as a whole rather than collections of
discrete audio events pre-extracted. Their process includes
finding the decomposition of a signal from a dictionary of
atoms, which would yield the best set of functions to form
an approximate representation. More recently, Mäkinen et
al. [33] propose an evolutionary feature synthesis technique
to enhance common audio descriptors by using multi-
dimensional particle swarm optimization to search for the
optimal feature synthesis parameters. These works forward
the auditory scene research; however, not all the unpredict-
able structures of auditory environments can be directly
discriminated by a low-level feature subset.
Context-aware approach is still at its early stage com-
paring with the acoustic feature selection methods to the
best of our knowledge. Generally, context in an auditory
scene can be considered as two different levels towards
incorporating hints for scene understanding. The first level
characterizes the correlations of certain audio events inside
the same acoustic scene such as gunfire and the accom-
panied cries in a war, or ball-hit and the corresponding
applause in a basketball game. Niessen et al. [34] model
knowledge and context in audio recognition by investigat-
ing the role of dynamic network model to improve auto-
matic audio identification and simultaneously reduce the
search space of low-level audio features. Heittola et al.
[35] present each audio context using a histogram of audio
events which are detected using a supervised classifier
based on annotated recordings. In their further work [36],
they provide context rule knowledge to better describe the
search space. For example, it will determine excluding the
footsteps class when the tested recording is from inside
a car. However, the contexts are modeled by three-state
left-to-right hidden Markov parameters and thereby still
face difficulties because of the great number of possible
event combinations and the transitions among them. The
second context-aware level describes more general infor-
mation about the surroundings around an audio device in
spite of the location, such as time [37], weather [38], run-
ning pace [39] and even user-dependent states like emotion
[40] or physiological state [37]. The advent of smart mobile
phones with rich sensing capabilities is making real-time
context information collecting and exploration a possibility,
and new auditory scene recognition systems are expected in
the next few years.
Both the two categories of methods have to consider
their dependency on selecting a proper learning algo-
rithm to obtain their recognition results. Mirikitani and
Nikolaev [41] recursively train recurrent neural networks
for improved time-series modeling. In [42], a set of key
audio effects are modeled with HMMs and a Bayesian
network-based approach is proposed to discover the high-
level semantics of an auditory context embedded in key
effect sequences. Recently, the combination of different
classifiers has been proved effective in environmental audio
understanding. For example, in [43], a hybrid SVM/k-
nearest neighbor (kNN) classifier is used for environmen-
tal audio classification based on MPEG-7 audio low-level
descriptors. Räsänen et al. [44] study the combination of
classifier output distributions using a number of different
classifiers instead of performing audio context fusion at a
feature level. Kinnunen et al. [45] choose MFCC param-
eterization and present an extensive comparison of six dif-
ferent classifiers of kNN, vector quantization (VQ), Gauss-
ian mixture model trained with both maximum likelihood
and maximum mutual information (GMM-MMI) criteria,
GMM supervector support vector machine and SVM with
generalized linear discriminant sequence for auditory
scene recognition. They find GMM-MMI and VQ classifi-
ers perform the best identification rates. On the other side,
multimodality algorithms such as probabilistic models to
integrate voice and visual identification cues gained from
microphones and cameras in a smart environment have also
been proposed [4, 46], which do help in automatic acoustic