Analysis of the Data Quality of Audio Features of Environmental Sounds

Abstract: In this paper we perform statistical data analysis of a broad set of state-of-the-art audio features and low-level MPEG-7 audio descriptors. The investigation comprises data analysis to reveal redundancies between state-of-the-art audio features and MPEG-7 audio descriptors. We introduce a novel measure to evaluate the information content of a descriptor in terms of variance. Statistical data analysis reveals the amount of variance contained in a feature. It enables identification of independent and redundant features. This approach assists in efficient selection of orthogonal features for content-based retrieval. We believe that a good feature should provide descriptions with high variance for the underlying data. Combinations of features should consist of decorrelated features in order to increase expressiveness of the descriptions. Although MPEG-7 is a popular and widely used standard for multimedia description, only few investigations do exist that address analysis of the data quality of low-level MPEG-7 descriptions.

Keywords: content-based multimedia retrieval, feature extraction, feature analysis, statistical data analysis, low-level MPEG-7 audio descriptors.

Categories: H.3.1, H.3.3, G.3

1 Introduction

In the last decades a huge number of features was developed for the analysis of audio content. One of the first application domains of audio analysis was speech recognition [Rabiner, 93]. With upcoming novel application areas the analysis of music and general purpose environmental sounds gained importance. Different research fields evolved, such as audio segmentation, music information retrieval (MIR), and environmental sound recognition (ESR). Each of these areas developed its specific description techniques (features). Currently, features are often employed in other domains than their original ones. A recent effort to standardize multimedia description tools led to the MPEG-7 standard. MPEG-7 is an ISO/IEC standard for multimedia content description [ISO/IEC, 02]. The standard defines low-level descriptions techniques (including audio) as well as high-level tools for multimedia processing.

The huge number of existing features makes the selection of the most appropriate feature set for a task difficult. Statistical data analysis can help in the identification of independent features. In this paper, we perform a quantitative analysis of the so called MPEG-7 low-level audio descriptors (LLDs) in the domain of environmental sounds. In the following we consider the descriptors as features and use the terms interchangeably. We compare MPEG-7 descriptors to a set of state-of-the-art audio features we previously analyzed in the domain of environmental sounds [Mitrovic, 06].

We investigate different description techniques by statistical data analysis in order to identify similarities and redundancies. Redundant features describe similar properties of the underlying data, while statistically independent features contain orthogonal information. The objective of feature selection is the combination of orthogonal features in order to maximize the amount of represented information. The method proposed in this paper supports the identification of independent and redundant features. Furthermore, we evaluate selected MPEG-7 high-level tools from this point of view. Additionally, we investigate the amount of information (entropy) contained in each feature. The information content of a feature is proportional to the variance of the feature values for a given dataset. This assumption is independent from the used similarity measure and classification technique. A discriminative feature must have high variance for a set of inhomogeneous sound samples. We derive a measure that represents the information contained in a feature with respect to its variance in order to evaluate the expressiveness of a feature.

The remainder of the paper is organized as follows. Section 2 gives background information on the MPEG-7 descriptors. In Section 3, we describe the approach for the data analysis. We present the structure of the experiments in Section 4. The results of the experiments are discussed in Section 5. We present some related work in Section 6. Finally we conclude the results in Section 7.

2 Background

The MPEG-7 Audio part specifies data structures and techniques for the description of audio content. It contains low-level descriptors (LLDs) as well as more sophisticated description techniques. In this section, we discuss the MPEG-7 LLDs relevant to this study.

2.1 Low-level MPEG-7 Audio Descriptors

The MPEG-7 Audio LLDs are a collection of low-level audio features that describe characteristic properties of sound such as harmonicity, sharpness, pitch, and timbre [Kim, 05]. The descriptors are applied to short frames of the signal and are either scalars or vectors per frame. Aggregation of descriptions of several frames to a description of an entire media object is not a normative part of the MPEG-7 standard. Some LLDs (TimbralSpectral descriptors) support a single-valued summarization for entire media objects. For other LLDs the standard proposes mathematical operations, such as minimum, maximum, mean, and variance for summarization. The MPEG-7 Audio LLDs are organized in the six groups: Basic, Basic Spectral, Spectral Basis, Signal Parameters, Timbral Temporal, Timbral Spectral.

In the following, we describe the MPEG-7 Audio LLDs together with their perceptual meaning and their application domain. A more detailed description can be found in [Kim, 05].

2.2 Basic

The LLDs in the Basic group primarily enable a short description of the shape of an audio waveform. The AudioWaveform (AW) descriptor represents the waveform envelope and is mainly intended for economic display of a waveform in an audio editor. AW comprises the minimum and maximum values of a framed signal. The AudioPower descriptor computes the average square of the waveform samples in a frame. It describes the power of the signal over time.

2.3 Basic Spectral

The LLDs in the BasicSpectral group describe basic properties of the spectrum of an audio signal. The AudioSpectrumEnvelope (ASE) descriptor represents the short-term power spectrum of a signal with a logarithmic frequency scale in several frequency bands. The logarithmic frequency scale aims at imitating properties of the human ear. The ASE descriptor is the basis for the computation of the other descriptors in the BasicSpectral group.

The AudioSpectrumCentroid (ASC) is the center of gravity of the spectrum calculated by ASE. The ASC descriptor indicates whether high or low frequencies dominate the spectrum of the signal. The AudioSpectrumSpread (ASS) descriptor represents the deviation of the power spectrum from its centroid. ASS enables separation of tonal sounds from noise-like sounds.

The fourth descriptor of the BasicSpectral group is AudioSpectrumFlatness (ASF). ASF describes the deviation of the spectrum of an audio signal from a flat shape. A flat spectrum indicates a noise-like or impulse-like signal. According to the MPEG-7 standard ASF is designed to perform fingerprinting, which requires robust matching between pairs of audio signals.

2.4 Spectral Basis

The SpectralBasis descriptors AudioSpectrumBasis (ASB) and AudioSpectrumProjection (ASP) are techniques for general-purpose sound recognition. ASB transforms the spectrum of a signal to a much lower-dimensional representation under certain statistical constraints. The ASB descriptor is based on the power spectrum, similarly to the ASE descriptor and provides a compact representation of a spectrum, while preserving a maximum amount of information.

The ASP is used together with the ASB descriptor. ASP takes a decibel-scaled spectrum as input and projects it against spectral basis functions, previously computed by ASB.

2.5 Signal Parameters

The SignalParameters group contains the AudioFundamentalFrequency (AFF) descriptor and the AudioHarmonicity (AH) descriptor. The AFF descriptor represents the fundamental frequency of a sound. AFF may be applicable to sound segmentation of speech and music. AH is a measure for the degree of harmonicity in a signal. The descriptor comprises of two components: harmonic ratio and upper limit of harmonicity. The harmonic ratio is the proportion of harmonic components in a signal. A purely harmonic signal has a harmonic ratio of "1," while the harmonic ratio of noise is "0." The upper limit of harmonicity specifies the frequency beyond which the audio signal has no more significant harmonic components.

2.6 Timbral Temporal

Timbral descriptors are usually employed in MIR. Timbre is a sound property that is independent of pitch and loudness. The LogAttackTime (LAT) characterizes the attack of a sound. The attack time is the time from the beginning of a sound signal to a point in time where its amplitude reaches a maximum. LAT is the logarithm of the attack time. The attack characterizes the beginning of a sound, which can be smooth or sudden. LAT may be employed for classification of musical instruments. The TemporalCentroid (TC) is the point in time where most of the signal energy is located.

2.7 Timbral Spectral

Harmonic peaks in a spectrum correspond to frequencies that are a multiple of the fundamental frequency. They are appropriate to describe the timbre of a signal. The TimbralSpectral descriptors rely on harmonic peak estimation by the fundamental frequency of the audio signal. The HarmonicSpectralCentroid (HSC) is the amplitude-weighted average of the harmonic peaks in a spectrum. The HarmonicSpectralSpread (HSS) descriptor is the amplitude-weighted deviation of the harmonic peaks from the HSC.

The HarmonicSpectralDeviation (HSD) is the deviation of the harmonic peaks from the spectral envelope. The spectral envelope is the mean over a few neighboring harmonic peaks. HarmonicSpectralVariation (HSV) refers to the correlation of harmonic peaks in adjacent frames. The fifth TimbralSpectral descriptor is SpectralCentroid (SC), which is the power-weighted average of the frequencies in the power spectrum.

The timbre descriptors are usually applied to music information retrieval in which timbre plays an important role. We investigate the applicability of timbre descriptors in the domain of environmental sounds. The descriptors are expected to yield average results in the experiments.

3 Approach

The quality of a feature may be measured by the amount of variance of its numeric values. A good feature should provide descriptions with high variance for the underlying data. Statistical data analysis reveals the amount of variance of a feature.

The analysis steps are as follows: Firstly features are extracted from the raw sound samples. Features based on audio frames are aggregated over time in order to obtain a description for the entire sound sample. The feature vectors of all sample files are combined into a matrix. This matrix is the basis for the statistical data analysis of the features.

Data analysis starts with a dimension reduction via Principal Components Analysis (PCA). The PCA decorrelates the second statistical moments (variances) of the feature data. We select only the principal components (PCs) with an Eigenvalue greater or equal than "1" for data analysis.

PCs with lower Eigenvalues are not considered since they explain less variance than the original data. From the PCA we yield the factor loading matrix that describes the influence of PCs on the particular feature components and vice versa. The factor loading matrix has the PCs in its columns (ordered by descending Eigenvalues) and the features form the rows of the matrix. Each entry (factor loading) represents the influence of a feature on a PC. The factor loadings are in the interval [-1, 1]. A high absolute factor loading indicates a high degree of correlation between a feature and a PC. Features that load the same PCs are correlated. A Varimax rotation simplifies the interpretation of factor loadings by maximizing their variances.

We investigate the expressiveness of the features using entropy. Furthermore, we derive a novel measure for expressiveness based on the factor loading matrix, as described in the following section.

3.1 A novel measure for expressiveness of features

In order to quantify the information contained in a feature we sum the absolute factor loadings weighted by the corresponding percentage of explained overall variance. The result is normalized by the sum of variances of all PCs. We call this measure Weighted Average LoaDing Indicator (WALDI). It is computed as defined in Equation 1.

(1)

Where f_j is the j-th feature from a set of F features and the c_i are the C principal components. The variance of the i-th PC is denoted by σ²(c_i). The factor loading matrix L is a R^CxF matrix with C columns and F rows.

The WALDI of a feature is a measure of its information content in terms of orthogonal variances in the data. This value is proportional to the expressiveness of a feature. However, it does not contain information about redundancies among different features. This information can be derived from the factor loading matrix.

We build a graph based on the WALDI (see Figure 1). Peaks in the WALDI graph indicate high expressiveness, which may have two reasons: either the corresponding feature loads a large number of less important PCs or it highly loads the few most important PCs.

Additionally, we compare WALDI with the information entropy, introduced by Shannon and Weaver [Shannon, 49]. We determine the entropy as defined in Equation 2.

(2)

We consider features as random variables and quantize the observed values of the features into n=256 bins. Then we compute the probability p(i) of the i-th bin for a feature f. For n=256 the unit of entropy is bit per byte. The entropy is a measure for the information content of a variable. The uncertainty of a random variable is proportional to its entropy. A powerful feature should have high entropy. We discuss the entropy of MPEG-7 features in Section 5.1.

4 Experiments

The investigations in this paper employ a database containing of 940 sound samples from 9 classes of environmental sounds. The sounds can be categorized into noise-like and tonal sounds. This distinction is of interest since several MPEG - 7 descriptors model these properties. Table 1 summarizes the 9 different classes, the respective number of sound samples and predominant sound characteristics.

The audio data in the experiments are sampled at 11025 Hz and quantized to 8 bits. MPEG-7 descriptors are computed with the LLD extractor provided by the TU Berlin [TUBerlin, 06]. The investigations include 12 previously analyzed audio descriptors from different application domains listed in Table 2 [Mitrovic, 06]. The MPEG-7 descriptors are summarized in Table 3.

TCombination of MPEG-7 and non-MPEG-7 descriptors yields a 391-dimensional feature vector. Frame-based features are summarized by statistical moments (mean and variance) in order to obtain descriptions of entire media objects.

5 Results

Quantitative data analysis discloses the data quality of numerical features. The basis of the investigation is the factor loading matrix that shows the mapping of features to PCs. In the first step of the analysis we investigate the expressiveness of features.

5.1 Information content analysis

We evaluate the audio descriptors based on their expressiveness and compute therefore the WALDI for the descriptor components (see Section 3.1 for details). We average over all WALDIs of the descriptor components to gain a more compact representation. That is equivalent to the average amount of information contained in each descriptor component. Figure 1 depicts the resulting graph.

The BasicSpectral descriptors ASS, ASC, and ASF yield high WALDIs, because these descriptors have high loadings for the first few and most important PCs. In contrast to this, the ASB descriptor has a small WALDI value. The reason for this is that the components of ASB do not load any of the important PCs. The factor loading matrix reveals a similar situation for ASE. This can as well be observed in the WALDI graph (see Figure 1). The timbral descriptors yield average values for the environmental sounds in the experiments. This is an unexpectedly good result since they mainly represent characteristics of musical sounds.

We believe that high entropy is a necessary property of a good feature. In the following, we compare the results obtained from the WALDI graph with the entropy of the descriptors. Figure 2 illustrates the average entropy of all components of a descriptor.

AH is the descriptor with the highest entropy (7.2 bit per byte), followed by the BasicSpectral descriptors, TC, and SC. Generally, the entropy of the LLDs is high with an average of 6 bit per byte. We observe that the MPEG-7 audio LLDs have higher entropy than the visual descriptors of the standard [Eidenberger, 04].

A comparison to WALDI and entropy shows that both measures correlate to a certain degree. Features with high WALDI tend to have high entropy. This is evident for the BasicSpectral descriptors (ASF, ASS, and ASC). Analogously, AP and ASE have low values for both measures. The correlation between WALDI and entropy shows that WALDI is a valid measure for information content and expressiveness.

5.2 Redundancy analysis

Identification of redundancies between features is an important pre-processing step of feature selection. The factor loading matrix is the basis for redundancy analysis. High absolute loadings indicate a high degree of correlation with the corresponding PC. Features that load the same PCs are dependent and contain redundant information.

The total redundancy of MPEG-7 descriptors is low, compared to the non-MPEG-7 features. In order to explain 85% of the overall variance in the data the non-MPEG-7 descriptors require 40 out of 228 PCs (~17.5%). MPEG-7 descriptors require 56 out of 163 PCs (~34.4%) to achieve the same result. Consequently, the set of MPEG-7 descriptors is less redundant.

In the following, we discuss the quality of the MPEG-7 Audio LLDs in more detail. The MPEG-7 descriptors of the Basic group (AP and AW) are highly correlated. The reason is that the computations of these two descriptors are similar. The descriptors represent related properties of the audio waveform. As expected, the expressiveness of these descriptors is limited. Both descriptors do not describe independent information with respect to the other features in the investigation. Most information explained by AP and AW is also contained in the Sone feature that measures the loudness of a signal.

The components of ASE are independent to a high degree. They load several less important PCs. However, ASE correlates with some Sone components. ASC, ASS, and ASF are highly correlated descriptors. The dependency of ASC and ASS is difficult to interpret since the descriptors describe different statistical moments. This may be a side effect introduced by the underlying data. In contrast to this, the redundancy of ASS and ASF is easier to explain since both descriptors model the same sound characteristics (noise-likeness and tonality). ASC, ASS, and ASF are highly redundant with several non-MPEG-7 features such as PLP, MFCC, and Sone.

The components of the SpectralBasis descriptor ASB are decorrelated. This is due to the SVD that yields orthogonal basis functions. The ASP descriptor bases on ASB. Thus, it correlates with ASB to some degree. ASB and ASP should be applied in conjunction according to the MPEG-7 standard. The factor loading matrix reveals the independence of ASB and ASP from all other features in this survey. These properties qualify ASB and ASP for combination with other features. Due to the similar computation of ASB and ASP, the two descriptors have comparable loadings in the factor loading matrix.

AH describes the harmonic structure of a signal. In the experiments AH is highly correlated with ASC, ASS and ASF because these descriptors depend on harmonic properties as well. Pitch, ZCR, and AD encode similar information as AH. A group of dependent MPEG-7 descriptors are AFF, ASC, SC, and HSC. They are correlated since they are all sensitive to the fundamental frequency of a signal. Consequently, the ZCR, which approximates the fundamental frequency, highly correlates with these features.

The timbral temporal descriptors LAT and TC are highly correlated with SF for the data in the experiments, even though they are computed entirely different. The reason for the high correlation is that these features do not model the characteristic of the environmental sounds very well. This is evident in the analysis of the expressiveness of these features (see Section 5.1). According to the WALDI graph in Figure 1, TC and LAT have average and low expressiveness, respectively.

This is due to the fact that the shape of environmental sounds is generally not structured in attack, decay, sustain and release as is the case for musical sounds. The expressiveness of SF with a WALDI of 0.7 is low as well because of the complexity and noise-like structure of the environmental sounds. However, LAT, TC, and SF describe information that is not captured by the other features in the study. That qualifies them for combination with other features.

The timbral spectral descriptors are completely redundant for the environmental sounds in the experiments. However, they describe unique information that is not captured by any other feature in the experiment. Since environmental sounds contain only little timbral characteristics the expressiveness of the timbral features is limited. Nevertheless timbral descriptors may be applicable to separate certain sound classes, such as bird sounds from other environmental sounds.

5.3 Summary

The following major insights can be derived from the analysis:

1. Most of the groups of MPEG-7 descriptors contain highly correlated components (Basic, BasicSpectral, SpectralBasis, TimbralTemporal, and TimbralSpectral). One reason may be the characteristics of the environmental sounds, which contain only little timbre and harmonicity. Another reason is the similarity of the computations of the descriptors in particular groups.
2. The descriptors in the SignalParameters group are decorrelated to a higher degree than the components of the other groups. This is because AH and AFF describe different signal properties.
3. The different descriptor groups are mostly independent from each other. The exceptions are the Basic group and the BasicSpectral group, which are highly correlated. The reason is that the two groups describe a signal waveform in time and frequency domain without further processing. Since the time and frequency representations of a signal are equivalent, the descriptions correlate. Two other correlated groups are BasicSpectral and SignalParameters. Both have correlated components such as AH and ASF, which describe the tonality of a sound. Further correlated components are ASC and AFF, which both heavily depend on the fundamental frequency.
4. Descriptors of the SpectralBasis and TimbralSpectral groups are independent from all other features including the non-MPEG-7 ones. Hence, they complement all possible feature combinations and are potential candidates for feature selection.
5. Several non-MPEG-7 features correlate with MPEG-7 descriptors. The Sone feature is redundant with the entire Basic and BasicSpectral groups. Pitch and PLP highly correlate with the BasicSpectral group. Further, several popular speech recognition features such as MFCC, BFCC, RASTA-PLP, Sone, Pitch, and PLP encode the information captured by ASF. The reason may be that all these features describe properties necessary for audio fingerprinting.

5.4 MPEG-7 high-level tools

The MPEG-7 high-level tools contain application-specific description schemes that build upon LLDs. These description schemes are AudioSignature, Timbre, and SoundModel. Other high-level descriptions do not consist of the LLDs discussed above and are therefore out of scope for the data analysis in this study.

The AudioSignature description scheme contains statistical summarizations of the ASF low-level descriptor. Its addressed application area is fingerprinting. Data analysis reveals that the components, which represent the ASF in different frequency bands, are highly redundant. Hence, a subset of ASF components may be sufficient for retrieval applications. Furthermore, it satisfies the requirements for fingerprinting since it contains a significant amount of information (see the WALDI graph in Figure 1).

Timbre is the second description scheme considered. It contains LAT, HSC, HSD, HSS, and HSV for harmonic instrument identification and SC, TC, and LAT for percussive instrument identification. In the experiments, we evaluate the quality of the timbral descriptors to prove whether timbre is a discriminative characteristic of environmental sounds. Experiments show that the descriptors for percussive instruments model environmental sounds better. The harmonic instrument descriptors are highly redundant for ES. However, they describe unique information with respect to the other descriptors in the experiments. Thus, only a subset of components of the Timbre description scheme may be feasible for ESR.

The SoundModel descriptor scheme is the third high-level tool in this investigation. It consists of ASB and ASP and addresses environmental sound recognition. While the components of ASB and ASP are highly decorrelated, their expressiveness is limited because they have only mediocre influence on the important PCs in the factor loading matrix. It may therefore be advantageous to combine the descriptors of SoundModel with other more expressive LLDs such as ASF.

6 Related Work

The MPEG-7 Audio standard provides a large set of low-level audio descriptors [ISO/IEC, 02]. MPEG-7 audio descriptors are part of many state-of-the-art audio retrieval systems [Kim, 04], [Benetos, 05], and [Xiong, 03]. MPEG-7 audio descriptors are applicable to different types of sound, such as music, speech and environmental sounds.

There are only few studies that address quantitative data analysis of content-based descriptions. We investigate low-level MPEG-7 visual descriptors from a statistical point of view in [Eidenberger, 04]. We employ statistical moments, factor analysis, and cluster analysis in the study. The investigation reveals that the visual descriptors are highly dependent on each other.

Most investigations apply features to a specific problem without a preceding data analysis. Such studies evaluate the quality of features empirically by performance measures such as recall and precision. A popular application domain is music information retrieval. In [Benetos, 05] the authors combine MPEG-7 descriptors with other common audio features for musical instrument classification. The investigation comprises of MPEG-7 BasicSpectral descriptors (ASE, ASC, ASS, and ASF) and SpectralBasis descriptors (ASB and ASP). Furthermore, SC and AH descriptors are incorporated. Additionally, the authors employ non-MPEG-7 features such as MFCCs, ZCR, and Spectral Rolloff Frequency [Kim, 05]. After feature extraction different classifiers (HMM, GMM and non-negative matrix factorization) predict the class membership in terms of six different classes of instruments.

MPEG-7 audio descriptors for instrument characterization are surveyed in [Casey, 01] and [Peeters, 00] as well. The authors investigate the quality of high-level descriptors (HarmonicInstrumentType and PercussiveInstrumentType) for the distinction of instrument sounds. They show that the combination of the low-level descriptors contained in the high-level description schemes enable successful similarity matching in a musical sounds database. The authors of [Allamanche, 01] present an MPEG-7 supported audio identification system that is robust with respect to common types of signal alterations. A flatness measure, similar to the MPEG-7 ASF descriptor is employed for similarity matching in a database of songs.

There are multiple investigations that deal with more general sounds such as environmental sounds. In [Casey, 01] the author presents the MPEG-7 sound recognition tools for general purpose sound recognition. The system employs the ASB and ASP descriptors. The author discusses the corresponding high-level description scheme (SoundModel) which includes continuous HMMs for classification. The authors of [Kim, 05], and [Kim, 04] extensively investigate the quality of the MPEG-7 sound recognition tools in different application domains. The authors perform speaker recognition and audio segmentation of speech and non-speech segments. Furthermore, they perform general sound classification of selected environmental sounds such as "Dog", "Bell", "Water", and "Baby". They compare the performance of MPEG-7 techniques with the traditional MFCC approach, originally developed for speech recognition. The MPEG-7 system employs ASB and ASP descriptors to represent the audio samples. Classification is performed by continuous HMMs. The investigations show that MPEG-7 descriptors perform comparably to MFCCs. However, MFCCs outperform ASB and ASP in some applications. A similar investigation is presented in [Xiong, 03]. The authors compare the widely used MFCC audio features to the low-level MPEG-7 descriptors designed for audio retrieval. Classification is performed with two types of HMMs (Maximum Likelihood HMM and Entropic Prior HMM). Again, MPEG-7 descriptors perform comparably to MFCCs. In [Wang, 03] the authors present an MPEG-7-based retrieval system for environmental sounds. The system applies ASC, ASS and ASF to the description of audio samples. The samples are organized in classes, such as "doorbell," "laughing," "knock", and "dog barks". One HMM represents one particular class.

The MPEG-7 standard defines a set of high-level description schemes in addition to the low-level description tools. An overview is given in [Quackenbush, 01]. The authors present MPEG-7 applications, such as Query by Humming, Audio Editing and application specific high-level description schemes (e.g. MusicalInstrumentTimbre, SpokenContent, and Melody).

7 Conclusions

In this paper we perform an extensive statistical data analysis of low-level MPEG-7 descriptions in the domain of environmental sounds. We analyze correlations of MPEG-7 descriptors with other state-of-the-art audio features. Statistical data analysis allows us to identify features that are complementary to MPEG-7 LLDs. We extend data analysis by introducing a novel measure for information content.

The objective of the experiments is the analysis of the correlations in a large set of features. For this purpose, we employ Principal Components Analysis, which reveals low redundancy between most of the MPEG-7 descriptor groups. However, there is high redundancy within some groups of descriptors such as the BasicSpectral group and the TimbralSpectral group. Redundant features capture similar properties of the media objects and should not be used in conjunction. We discuss the quality of MPEG-7 high-level description tools based on the results of the analysis of the LLDs. The entropy of the LLDs is generally high (in average 6 bit per byte) which is significantly more than for the visual descriptors we surveyed in [Eidenberger, 04].

Acknowledgements

The authors would like to express their gratitude to Professor Christian Breitenender for his guidance. We want to thank Professor Sikora from the Technical University of Berlin for providing us with the MPEG-7 LLD extractor.

References

[Allamanche, 01] E. Allamanche, et al, Content-based Identification of Audio Material Using MPEG-7 Low-level Description, In Proc. of the Int. Conf. on Music Information Retrieval, 2001, 197-204

[Benetos, 05] E. Benetos, et al, Comparison of Subspace Analysis-Based and Statistical Model-Based Algorithms for Musical Instrument Classification, 2nd Workshop on Immersive Communication and Broadcast Systems, 2005

[Casey, 01] M. Casey, MPEG-7 sound recognition tools, In IEEE Trans. on Circuits and Systems for Video Technology, vol 11, 2001, 737-747

[Eidenberger, 04] Eidenberger, H. Statistical analysis of content-based MPEG-7 descriptors for image retrieval, Multimedia Systems, vol. 10, 2004, 84-97

[ISO/IEC, 02] ISO/IEC 15938, Information Technology ? Multimedia Content Description Interface, First Edition, 2002

[Kim, 05] H. Kim, N. Moreau, T. Sikora, MPEG-7 audio and beyond, West Sussex: Wiley, 2005

[Kim, 04] H. Kim, T. Sikora, Comparison of MPEG-7 audio spectrum projection features and MFCC applied to speaker recognition, sound classification and audio segmentation, In Proc. of the Int. Conf. on Acoustics, Speech, and Signal Processing, vol 5, 2004, 925-928

[Mitrovic, 06] D. Mitrovic, M.Zeppelzauer, H. Eidenberger, Towards an Optimal Feature Set for Environmental Sound Recognition, Technical Report TR-188-2-2006-03, 2006, http://www.ims.tuwien.ac.at/publication_master.php

[Peeters, 00] G. Peeters, S. McAdams, P. Herrera, Instrument sound description in the context of MPEG-7, In Proc. of the 2000 Int. Computer Music Conf., 2000

[Quackenbush, 01] S. Quackenbush, A. Lindsay, Overview of MPEG-7 Audio, In IEEE Trans. on Circuits Systems for Video Technology, vol 11, 2001, 725-729

[Rabiner, 93] L. Rabiner, B. Juang, Fundamentals of speech recognition, New York: Prentice-Hall, 1993

[Shannon, 49] C. Shannon, W. Weaver, The Mathematical Theory of Communication, Urbana University of Illinois Press, 1949

[TUBerlin, 06] MPEG-7 Audio Analyzer Low Level Descriptors Extractor, http://mpeg7lld.nue.tu-berlin.de, 2006

[Wang, 03] J-F. Wang, et al, Home environmental sound recognition based on MPEG-7 features, In Proc. of the 46th IEEE Int. Midwest Symposium on Circuits and Systems, vol 2, 2003, 682-685.

[Xiong, 03] Z. Xiong, et al, Comparing MFCC and MPEG-7 audio features for feature extraction, Maximum Likelihood HMM and Entropic Prior HMM for sports audio classification, In Proc. of the Int. Conf. on Multimedia and Expo, vol. 3, 2003, 397-400