Using Generic Summarization to improve Music Information Retrieval Tasks

Many Music Information Retrieval (MIR) tasks process only a segment of the whole music signal, in order to satisfy processing time constraints. This practice may lead to decreasing performance since the most important information for the task may not be in those processed segments. In this paper, we leverage generic summarization algorithms, previously applied to text and speech summarization, to summarize items in music datasets. These algorithms build summaries, that are both concise and diverse, by selecting appropriate segments from the input signal which makes them good candidates to summarize music as well. We evaluate the summarization process on binary and multiclass music genre classification tasks, by comparing the performance obtained using summarized datasets against the performances obtained using continuous segments (which is the traditional method used for addressing the previously mentioned time constraints) and full songs of the same original dataset. We show that GRASSHOPPER, LexRank, LSA, MMR, and a Support Sets-based Centrality model improve classification performance when compared to selected 30-second baselines. We also show that these summarized datasets lead to a classification performance whose difference is not statistically significant from using full songs. Furthermore, we make an argument stating the advantages of sharing summarized datasets for future MIR research.


I. INTRODUCTION
Music summarization has been the subject of research for at least a decade and many algorithms that address this problem, mainly for popular music, have been published in the past [1]- [8]. However, those algorithms focus on producing human consumption-oriented summaries, i.e., summaries that will be listened to by people motivated by the need to quickly get the gist of the whole song without having to listen to all of it. This type of summarization entails extra requirements besides conciseness and diversity (non-redundancy), such as clarity and coherence, so that people can enjoy listening to them.
Generic summarization algorithms, however, focus on extracting concise and diverse summaries and have been successfully applied in text and speech summarization [9]- [13]. Their application, in music, for human consumption-oriented purposes is not ideal, for they will select and concatenate the most relevant and diverse information (according to each algorithm's definition of relevance and diversity) without taking into account whether the output is enjoyable for people or not. This is usually reflected, for instance, on discontinuities or irregularities in beat synchronization in the resulting summaries.
We focus on improving the performance of tasks recognized as important by the MIR community, e.g. music genre classification, through summarization, as opposed to considering music summaries as the product to be consumed by people. Thus, we can ignore some of the requirements of previous music summarization efforts, which usually try to model the musical structure of the pieces being summarized, possibly using musical knowledge. Although human-related aspects of music summarization are important in general, they are beyond the focus of this paper. We claim that, for MIR tasks benefiting from summaries, it is sufficient to consider the most relevant parts of the signal, according to its features. In particular, summarizers do not need to take into account song structure or human perception of music. Our rationale is that summaries contain more relevant and less redundant information, thus improving the performance of tasks that rely on processing just a portion of the whole signal, leading to faster processing, less space usage, and efficient use of bandwidth.
We use GRASSHOPPER [12], LexRank [10], LSA [11], MMR [9], and Support Sets [13] to summarize music for automatic (instead of human) consumption. To evaluate the effects of summarization, we assess the performance of binary and 5-class music genre classification, when considering song summaries against continuous clips (taken from the beginning, middle, and end of the songs) and against the whole songs. We show that all of these algorithms improve classification performance and are statistically not significantly different from using the whole songs. These results complement and solidify previous work evaluated on a binary Fado classifier [14].
The article is organized as follows: section II reviews related work on music-specific summarization. Section III reviews the generic summarization algorithms we experimented with: GRASSHOPPER (section III-A), LexRank (section III-B), LSA (section III-C), MMR (section III-D), and Support Setsbased Centrality (section III-E). Section IV details the experiments we performed for each algorithm and introduces the classifier. Sections V and VI report our classification results for the binary and multiclass classification scenarios, respectively. Section VII discusses the results and Section VIII concludes this paper with some remarks and future work.
II. MUSIC SUMMARIZATION Current algorithms for music summarization were developed to extract an enjoyable summary so that people can listen to it clearly and coherently. In contrast, our approach considers summaries exclusively for automatic consumption.
Human-oriented music summarization starts by structurally segmenting songs and selecting meaningful segments to in-clude in the summary. The assumption is that songs are represented as label sequences where each label represents a different part of the song (e.g., ABABCA where A is the chorus, B the verse, and C the bridge). In [1], segmentation is achieved by using a Hidden Markov Model (HMM) to detect key changes between frames and Dynamic Time Warping (DTW) to detect repeating structure. In [2], a Gaussiantempered "checkerboard" kernel is correlated along the main diagonal of the song's self-similarity matrix, outputting segment boundaries. Then, a segment-indexed matrix, containing the similarity between detected segments, is built. Singular Value Decomposition (SVD) is applied to find its rank-K approximation. Segments are, then, clustered to output the song's structure. In [3], [4], a similarity matrix is built and analyzed for fast changes, outputting segment boundaries; segments are clustered to output the "middle states"; an HMM is applied to these states, producing the final segmentation. Then, various strategies are considered to select the appropriate segments.
In [5], a modification of the Kullback-Leibler (KL) divergence is used to group and label similar segments. The summary consists of the longest sequence of segments belonging to the same cluster. In [6] and [7], Average Similarity is used to extract a thumbnail L seconds long that is the most similar to the whole piece. It starts by calculating a similarity matrix through computing frame-wise similarities. Then, it calculates an aggregated similarity measure, for each possible starting frame, of the L-second segment with the whole song and picks the one that maximizes it as the summary. Another method for this task, Maximum Filtered Correlation [8], starts by building a similarity matrix and then a filtered time-lag matrix, embedding the similarity between extended segments separated by a constant lag. The starting frame of the summary corresponds to the index that maximizes the filtered time-lag matrix. In [15], music is classified as pure or vocal, in order to perform type-specific feature extraction. The summary, created from three to five seconds subsummaries (built from frame clusters), takes into account musicological and psychological aspects, by differentiating between types of music based on feature selection and specific duration. This promotes human enjoyment when listening to the summary. Since these summaries were targeted to people, they were evaluated by people.
In [16], music datasets are summarized into a codebookbased audio feature representation, to efficiently retrieve songs in a query-by-tag and query-by-example fashion. An initial dataset is discretized, creating a dictionary of k basis vectors. Then, for each query song, the audio signal is quantized, according to the pre-computed dictionary, mapping the audio signal into a histogram of basis vectors. These histograms are used to compute music similarity. This type of summarization allows for efficient retrieval of music but is limited to the features which are initially chosen. Our focus is on audio signal summaries, which are suitable for any audio feature extraction, instead of proxy representations for audio features.

III. GENERIC SUMMARIZATION
Applying generic summarization to music implies song segmentation into musical words and sentences. Since we do not take into account human-related aspects of music perception, we can segment songs according to an arbitrarily fixed size. This differs from structural segmentation in that it does not take into account human perception of musical structure and does not create meaningful segments. Nevertheless, it still allows us to look at the variability and repetition of the signal and use them to find its most important parts. Furthermore, since it is not aimed at human consumption, the generated summaries are less liable to violate the copyrights of the original songs. This facilitates the sharing of datasets (using the signal itself, instead of specific features extracted from it) for MIR research efforts. In the following sections, we review the generic summarization algorithms we evaluated.

A. GRASSHOPPER
The Graph Random-walk with Absorbing StateS that HOPs among PEaks for Ranking (GRASSHOPPER) [12] was applied to text summarization and social network analysis, focusing on improving ranking diversity. It takes an n×n matrix W representing a graph where each sentence is a vertex and each edge has weight w ij corresponding to the similarity between sentences i and j; and a probability distribution r encoding prior ranking. First, W is row-normalized: O ij =w ij / n k=1 w ik . Then, P =λO+(1−λ)1r T is built, incorporating the user-supplied prior ranking r (1 is an all-1 vector, 1r T is the outer product, and λ is a balancing factor). The first ranking state g 1 = argmax n i=1 π i is found by taking the state with the largest stationary probability (π=P T π is the stationary distribution of P ). Each time a state is extracted, it is converted into an absorbing state to penalize states similar to it. The rest of the states are iteratively selected according to the expected number of visits to each state, instead of considering the stationary probability. If G is the set of items ranked so far, states are turned into absorbing states by setting P gg =1 and P gi =0, ∀i =g. If items are arranged so that ranked ones are listed before unranked ones, P can be written as follows: I G is the identity matrix on G. R and Q are rows of unranked items. N =(I−Q) −1 is the expected number of visits to state j starting from state i (N ij ). The expected number of visits to state j, v j , is given by v=(N T 1)/(n−|G|) and the next item

B. LexRank
LexRank [10] relies on the similarity (e.g. cosine) between sentence pairs (usually, tf-idf vectors). First, all sentences are compared to each other. Then, a graph is built where each sentence is a vertex and edges are created between every sentence according to their pairwise similarity (above a threshold). LexRank can be used with both weighted (eq. 2) and unweighted (eq. 4) edges. Then, each vertex score is iteratively computed. In eq. 2 through 4, d is a damping factor to guarantee convergence; N is the number of vertices; S (V i ) is the score of vertex i; and D (V i ) is the degree of i. Summaries are built by taking the highest ranked sentences.
In LexRank, sentences recommend each other: sentences similar to many others will get high scores. Scores are also determined by the score of the recommending sentences.
C. Latent Semantic Analysis (LSA) LSA was first applied in text summarization in [17]. SVD is used to reduce the dimensionality of an original matrix representation of the text. LSA-based summarizers start by building a T terms by N sentences matrix A. Each element of A, a ij =L ij G i , has a local (L ij ) and a global (G i ) weight. L ij is a function of term frequency in a specific sentence and G i is a function of the number of sentences that contain a specific term. Usually, a ij are tf-idf scores. The result of applying the SVD to A is A=U ΣV T , where U (T ×N matrix) are the left singular vectors; Σ (N ×N diagonal matrix) contains the singular values in descending order; and V T (N ×N matrix) are the right singular vectors. Singular values determine topic relevance: each latent dimension corresponds to a topic. The rank-K approximation considers the first K columns of U , the K×K sub-matrix of Σ, and the first K rows of V T . Relevant sentences are the ones corresponding to the indices of the highest values for each right singular vector. This approach has two limitations [18]: by selecting K sentences for the summary, less significant sentences tend to be extracted when K increases; and, sentences with high values in several topics, but never the highest, will never be included in the summary. To account for these effects, a sentence score was introduced and K is chosen so that the K th singular value does not fall under half of the highest singular value:

D. Maximal Marginal Relevance (MMR)
Sentence selection in MMR [9] is done according to their relevance and diversity against previously selected sentences, in order to output low-redundancy summaries. MMR is a query-based method that has been used in speech summarization [19], [20]. It is also possible to produce generic summaries by taking the centroid vector of all the sentences as the query.
MMR uses λSim 1 (S i , Q) − (1−λ) max Sj Sim 2 (S i , S j ) to select sentences. Sim 1 and Sim 2 are similarity metrics (e.g. cosine); S i and S j are unselected and previously selected sentences, respectively; Q is the query, and λ balances relevance and diversity. Sentences can be represented as tf-idf vectors.

E. Support Sets-based Centrality
This method was first applied in text and speech summarization [13]. Centrality is based on sets of sentences that are similar to a given sentence (support sets): S i ={s∈I : Sim (s, p i ) > i ∧s =p i }. Support sets are estimated for every sentence. Sentences frequent in most support sets are selected: argmax s∈∪ n i=1 Si |{S i : s∈S i }|. This is similar to unweighted LexRank (section III-B), except that support sets allow a different threshold for each sentence ( i ) and their underlying representation is directed, i.e., each sentence only recommends its most semantically related sentences. The thresholds can be heuristically determined. [13], among others, uses a passage order heuristic which clusters all passages into two clusters, according to their distance to each cluster's centroid. The first and second clusters are initialized with the first and second passages, respectively, and sentences are assigned to clusters, one by one, according to their original order. The cluster that contains the most similar passage to the passage associated with the support set under construction is selected as the support set. Several metrics were tested for defining semantic relatedness (e.g. Minkowski distance, cosine).

IV. EXPERIMENTS
We evaluated generic summarization by assessing its impact on binary and multiclass music genre classification. These tasks consist of classifying songs based on a scheme (e.g. artist, genre, or mood). Classification is deemed important by the MIR community and annual conferences addressing it are held, such as International Society for Music Information Retrieval (ISMIR), which comprises Music Information Retrieval Evaluation eXchange (MIREX) [21] for comparing state-of-the-art algorithms in a standardized setting. The best MIREX 2015 system [22] for the "Audio Mixed Popular Genre Classification" task uses Support Vector Machines (SVMs) for classifying music genre, based on spectral features.
We follow the same approach and our classification is also performed using SVMs [23]. Note that there are two different feature extraction steps. The first is done by the summarizers, every time a song is summarized. The summarizers output audio signal corresponding to the selected parts, to be used in the second step, i.e., when doing classification, where features are extracted from the full, segmented, and summarized datasets.

A. Classification Features
The features used by the SVM consist of a 38-dimensional vector per song, a concatenation of several statistics on features used in [24], describing the timbral texture of a music piece. It consists of the average of the first 20 Mel Frequency Cepstral Coefficients (MFCCs) concatenated with statistics (mean and variance) of 9 spectral features: centroid, spread, skewness, kurtosis, flux, rolloff, brightness, entropy, and flatness. These are computed over feature vectors extracted from 50ms frames without overlap. This set of features and a smaller set, solely composed of MFCC averages, were tested in the classification task. All music genres in our dataset are timbrically different from each other, making these sets good descriptors for classification.

B. Datasets
Our experimental datasets consist of a total of 1250 songs from 5 different genres: Bass, Fado, Hip hop, Trance, and Indie Rock. Bass music is a generic term referring to several specific styles of electronic music, such as Dubstep, Drum and Bass, Electro, and more. Although these differ in tempo, they share similar timbral characteristics, such as deep basslines and the "wobble" bass effect. Fado is a Portuguese music genre whose instrumentation consists of stringed instruments, such as the classical and the Portuguese guitars. Hip hop consists of drum rhythms (usually built with samples), the use of turntables and spoken lyrics. Indie Rock usually consists of guitar, drums, keyboard, and vocal sounds and was influenced by punk, psychedelia, post-punk, and country. Trance is an electronic music genre characterized by repeating melodic phrases and a musical form that builds up and down throughout a track. Each class is represented by 250 songs from several artists. The multiclass dataset contains all songs. Two binary datasets were also built from this data, in order to test our hypothesis on a wider range of classification setups: Bass vs. Fado and Bass vs. Trance, each containing the 500 corresponding songs.

C. Setup
10-fold cross-validation was used in all classification tasks. First, as baselines, we performed 3 classification experiments using 30s segments, from the beginning, middle, and end of each song. Then, we obtained another baseline by using the whole songs. The baselines were compared with the classification results from using 30s summaries for each parameter combination and algorithm. We did this for both binary datasets and then for the multiclass dataset.
Applying generic summarization algorithms to music requires additional steps. Since these algorithms operate on the discrete concepts of word and sentence, some preprocessing must be done to map the continuous frame representation obtained after feature extraction to a word/sentence representation. For each song being summarized, a vocabulary is created, through clustering the frames' feature vectors. mlpack's [25] implementation of the K-Means algorithm was used for this step (we experiment with some values for K and assess their impact on the results). After clustering, a vocabulary of musical words is obtained (each word is a frame cluster's centroid) and each frame is assigned its own cluster centroid, effectively mapping the frame feature vectors to vocabulary words. This transforms the real/continuous nature of each frame (when represented by a feature vector) to a discrete nature (when represented as a word from a vocabulary). Then, the song is segmented into fixed-size sentences (e.g., 5-word sentences). Since every sentence contains discrete words from a vocabulary, it is possible to represent each one as a vector of word occurrences/frequencies (depending on the weighting scheme) which is the exact representation used by generic summarization algorithms. Sentences were compared using the cosine distance. The parameters of all of these algorithms include: features, framing, vocabulary size (final number of clusters of the K-Means algorithm), weighting (e.g., tf-idf ), and sentence size (number of words per sentence).
For the multiclass dataset, we also ran experiments comparing human-oriented summarization against generic summarization. This translates into comparing Average Similarity summaries (for several durations) against 30-second generic summaries, as well as comparing structural against fixed-size sentences. We also compared the performance of generic summaries against the baselines for smaller summary durations.
Every algorithm was implemented in C++. We used: OpenS-MILE [26] for feature extraction, Armadillo [27] for matrix operations, Marsyas [28] for synthesizing the summaries, and the segmenter used in [29] for structural segmentation.
Our experiments covered the following parameter values (varying between algorithms): frame and hop size combinations of (0.25,0.125), (0.25,0.25), (0.5,0.25), (0.5,0.5), (1,0.5) and (1,1) (in seconds); vocabulary sizes of 25, 50, and 100 (words); sentence sizes of 5, 10, and 20 (words); "dampened" tf-idf (takes logarithm of tf instead of tf itself) and binary weighting schemes. As summarization features, we used MFCC vectors of sizes 12, 20, and 24. These features, used in several previous research efforts on music summarization in [1]- [7], describe the timbre of an acoustic signal. We also used a concatenation of MFCC vectors with the 9 spectral features enumerated in section IV-A. For MMR, we tried λ values of 0.5 and 0.7. Our LSA implementation also makes use of the sentence score and the topics cardinality selection heuristic described in section III-C.

V. RESULTS: BINARY TASKS
First, we analyze results on the binary datasets, Bass vs. Fado and Bass vs. Trance. The reason we chose these pairs was because we wanted to see summarization's impact on an easy to classify dataset (Bass and Fado are timbrically very different) and a more difficult one (Bass and Trance share many timbrical similarities due to their electronic and dancefloororiented nature). For all experiments, classifying using the 38dimensional features vector produced better results than using only 20 MFCCs, so we only present those results here. The best results are summarized in tables Ia, Ib, and Ic. The first thing we notice on the Bass vs. Fado task is that the middle sections are the best continuous sections and they do a good job at distinguishing Fado from other genres. Accuracy dropped just 2 percentage points (pp) against using full songs. However, the beginning sections' accuracy dropped by 5.8pp. All summarization algorithms fully recovered the accuracy lost by any continuous sections against using full songs, achieving the 100% full songs baseline. In this case, summarization helps classification in an already easy task. The λ value in MMR's setup was 0.7 and the passage order heuristic using the cosine similarity was used for calculating the support sets.
In the Bass vs. Trance task, the middle sections do a very poor job at describing and distinguishing these genres -they actually perform worse than the beginning or end sections. Actually, the worst sections in the Bass vs. Fado task were the best in this one and vice-versa. This means that choosing a continuous segment to extract features for classification purposes cannot be assumed to work equally well for every genre and dataset. All summarization algorithms, while not reaching the same performance as when using full songs, succeeded in improving classification performance against the continuous 30-second baselines. In this case, summarization is helping classification in a more difficult task. Again, MMR's λ value was set to 0.7 and the passage order heuristic using the cosine similarity was used for calculating the support sets.

VI. RESULTS: MULTICLASS TASKS
Since we are extrinsically evaluating summarization, analyzing its impact on music classification must go beyond simply comparing final classification accuracy for each scenario (as was done for binary classification). Here, we also look at the confusion matrices obtained from the classification scenarios, so that we can carefully look at the data (in this case, listen to the data) to understand what is happening when summarizing music this way and why it is improving the classification task's performance. Since our dataset consists of 250 songs per class, each confusion matrix row must sum to 250. Classes are identically sorted both in rows and columns, which means the ideal case is where we have a diagonal confusion matrix (all zeros, except for the diagonal elements, which should all be 250). Class name initials are shown to the left of the matrix and individual class accuracies are shown to the right.

A. Full songs
First, we look at the confusion matrix resulting from classifying full songs (table II). We can see that Fado, although there is some confusion between it and Indie Rock, is the most distinguishable genre within this group of genres which makes sense since timbrically it is very different from every other genre present in the dataset. Trance and Bass also achieve accuracies over 90%, although they also share some confusion which is explained by the fact that they both are Electronic music styles, thus sharing many timbral characteristics derived from the virtual instruments used to produce them. The classifier performs worse when classifying Hip hop and Indie Rock, achieving accuracies around 84% and confusing both genres in approximately 10% of the tracks. This can also be explained by the fact that both of those genres have strong vocals presence (in contrast with Bass and Trance). Although Fado also has an important vocal component, its instrumentation is very different from Hip hop and Indie Rock explaining why Fado did not get confused as much as they were with each other. The overall accuracy of this classification scenario is 89.84%.
We can think of these accuracies as how well these classification features (and SVM) can perform on these genres, given all the possible information about the tracks. Intuitively, removing information by, for instance, only extracting features from the beginning 30 seconds of the songs, will worsen the performance of the classifier because it will have incomplete data about each song, and thus, also incomplete data for modeling each class. Tables IIIa, IIIb, and IIIc show that to be true when using such a blind approach to summarize music (since extracting 30-second contiguous segments can also be interpreted as a naive summarization method). This process of extracting features from a dataset of segments is what is usually done when classifying music, since processing 30 seconds instead of the whole song saves processing time.  Table IIIa shows classification results when using only the 30 seconds from the beginning of the songs. Table IIId shows the comparison of the beginning sections against full songs. The classification accuracy is 77.52%, i.e., a 12.32pp drop when compared to using full songs. Bass accuracy dropped 19.6pp, due to increased confusion with both Hip hop and Indie Rock. Trance was also more confused with Indie Rock. This is easily explained by the fact that the first 30 seconds of most Bass or Trance songs correspond to the intro part. These intros are lower energy parts which may contain a relatively strong vocal presence and much fewer instrumentation than other more characteristic parts of the genres. These intros are much more similar to Hip hop and Indie Rock intros than when considering the whole songs, explaining why the classifier is confusing these classes more in this scenario. Thus, taking the beginning of the songs for classification is, in general, not a good summarization strategy.

B. Baseline segments
Tables IIIb and IIIe show classification results when using the middle 30 seconds of the songs and the comparison of those segments against full songs, respectively. The overall accuracy was 81.36%, i.e., an 8.48pp drop against the full songs baseline. This time, both Bass and Trance accuracies dropped 16.8pp and 20.4pp, respectively, getting confused with each other by the classifier. Having listened to the tracks that got confused this way, the conclusion is as expected: these middle segments correspond to what is called a breakdown section of the songs. These sections correspond to lower energy segments (though not as low as an intro) of the tracks which, again, are not the most characteristic parts of both these genres and, in the particular case of Bass vs. Trance, they are timbrically very similar due to their Electronic nature. A human listener would, probably, also be unable to distinguish between these two genres if listening only to these segments. Although, for 3 of the 5 genres, classification performance did not drop pronouncedly, it did so for 2 of them, which means that, in general, taking the middle sections of the songs for classification is also not a good segment selection strategy. Tables IIIc and IIIf show classification results when using the last 30 seconds of the songs and the comparison of those segments versus full songs, respectively. The end sections obtained an accuracy of 76.8%, i.e., a 13.04pp decrease when compared against full songs. Again, Bass was mainly misclassified as Hip hop and Indie Rock, and Trance was mainly misclassified as Bass. This is mostly due to the fact that the last 30 seconds correspond to the outro section of the songs which shares many similarities with the intro section. When considering Trance and Bass, the outro also shares characteristics with the breakdown sections. The fade repeat effect present in many songs' endings also increases this confusion. This means that taking the last 30 seconds of a song is also not a good segment selection strategy.

C. Baseline Assessment
Although, from the above experiments, it seems that taking the middle sections of the songs is better than taking the beginning or end, it is still not good enough, at least, not for all of the considered genres. The features used by the classifier are statistics (means and variances) of features extracted along the whole signal. Those features perform well when taking the whole signal as input, which means that, in order to obtain a similar performance, those statistics should be similar. That cannot be guaranteed when taking 30-second continuous clips because those 30 seconds may happen to belong to a single (and not distinctive enough) structural part of the song (such as intro, breakdown, and outro). If that is the case, then there is not sufficient diversity in the segment/summary to accurately represent the whole song. Moreover, some music genres can only be accurately distinguished by some of those structural parts: the best examples in this dataset are the Bass and Trance classes, which are much more accurately distinguished and represented by their drop sections. Therefore, we need to make better choices regarding what parts of the song should be included in the 30-second summaries to be classified.

D. GRASSHOPPER
Generic summarization algorithms define and detect relevance and diversity of the input signal, satisfying our need for a more informed way of selecting the most important parts to fit in 30-second summaries. The following tables show results demonstrating this claim. Tables IVa and IVb show classification results when using summaries extracted by GRASSHOP-PER. The specific parameter values used in this experiment were: (0.5,0.5) seconds framing, 25-word vocabulary, 10-word sentences, and binary weighting. The overall accuracy was 88.16%. As can be seen, GRASSHOPPER recovered most of what was lost by the middle sections, in terms of classification accuracy for each class. Since the middle sections performed so badly when distinguishing Bass and Trance, naturally, these summaries improved accuracies mostly for both these classes, with 14.0pp and 14.4pp increases, respectively. When listening to some of these summaries, the diversity included in them is clear: the algorithm is selecting sentences from several different structural parts of the songs. An overall improvement of 6.80pp was obtained this way. Note that, remarkably, these summaries did a better job than full songs at classifying Hip hop by 2.0pp. This means that, for some tasks, well summarized data can be even more discriminative of a topic (genre) than the original full data.

E. LexRank
Tables Va and Vb present the LexRank confusion matrix and its difference against the middle sections. The parameter values in this experiment were: (0.5,0.5) seconds framing, 25-word vocabulary, 5-word sentences and dampened tf-idf weighting. The overall accuracy was 88.40%. LexRank also greatly improved classification accuracy, when compared against the middle sections (7.04pp overall), namely, for Bass and Trance, with 15.6pp and 15.2pp increases, respectively. LexRank is clearly selecting diverse parts to include in the 30-second summaries, as we were able to conclude when listening to them. It is also interesting that the classifier performed better than with full songs, individually, for another class: Indie Rock's accuracy increased 1.6pp.

F. LSA
Tables VIa and VIb show the LSA confusion matrix of and the corresponding difference against the middle sections. The following parameter combination was used: (0.5,0.5) seconds framing, 25-word vocabulary, 10-word sentences, and binary weighting. Note that using a term frequency-based weighting on LSA, when applied to music, markedly worsens its performance. This is because noisy sentences in the songs tend to get a very high score on some latent topic, causing LSA to include them in the summaries. Moreover, when also considering inverse document frequency, the results are even worse, because those noisy terms usually appear in very few sentences. That is highly undesirable, since those sections do a very bad job at describing that song in any aspect. Using a binary weighting scheme alleviates that problem because all those noisy frames will get clustered into very few clusters/terms and only that term's presence (instead of frequency) gets counted into the sentences' vector representation. The overall accuracy for this combination was 88.32%, an improvement of 6.98pp against the middle sections. Bass and Trance were also the genres which benefited the most from this summarization, with accuracy increases of 12.8pp and 14.8pp, respectively, which can also be explained by the diversity present in the summaries. Indie Rock's individual accuracy improved, once again, against full songs, with an improvement of 2.8pp.

G. MMR
Tables VIIa and VIIb represent the confusion matrix for an MMR summarization setup and its difference against the middle sections. (0.5,0.5) seconds framing was used, along with a 50-word vocabulary, 10-word sentences, 0.7 λ value, and dampened tf-idf weighting. Note that, even though every other parameter setup (for the other algorithms) shown here uses 20 MFCCs as features, this one uses those same MFCCs concatenated with the 9 spectral features also used for classification (described in section IV-A). This is because MMR, unlike every other summarization algorithm, performed better using this set (instead of only using MFCCs as features). The overall accuracy was 88.80%, corresponding to an improvement of 7.44pp over the middle sections. Bass and Trance benefited the most from the summarization process, in classification performance, achieving improvements of 14.8pp and 16.4pp, respectively. This is also explained by the diversity produced by the summarizer.

H. Support Sets
Tables VIIIa and VIIIb show results obtained when classifying the dataset using summaries extracted by the Support Sets-based algorithm. The specific parameter setup of this experiment was: (0.5,0.5) seconds framing, 25-word vocabulary, 10-word sentences, dampened tf-idf weighting, and the passage order-based heuristic for creating the support sets [13] using the cosine similarity. The overall accuracy was 88.80%. Again, summarization recovered most of what was lost by the middle sections in terms of classification accuracy for each individual class, greatly influencing Bass and Trance, with 10.8pp and 16.8 pp increases, respectively. Listening to some of these summaries, we confirmed the diversity included in them that was clearly lacking in the middle sections. An overall improvement of 7.44pp was obtained this way. Remarkably, there were also improvements against full songs, namely, a 4.8pp improvement in Indie Rock.

I. Summary size experiments
To better evaluate the robustness of these methods, we ran experiments using decreasing summary sizes. For these experiments, no search for optimal parameter combinations was done: we used the ones that maximized classification accuracy for 30-second summaries. These are not necessarily the best parameters for smaller summary sizes but allow using the 30-second summaries as baselines. We ran these experiments for summary sizes of 5 to 25 seconds and report the results in table IX and Figure 1. Considering classification accuracy, every algorithm, except for MMR, outperforms the best 30-second baseline with just 5-second summaries. LSA and Support Sets, in particular, surpass the 87% accuracy mark using just 10-second summaries. Note that these experiments were not fine tuned.

J. Average Similarity
To obtain a human-oriented baseline, we summarized the dataset with Average Similarity (section II). This can be seen as an informed human-relevant way of selecting the best starting position of a contiguous segment. The parameter values used in this experiment were: (0.5,0.5) seconds framing, and the first 20 MFCCs as features. Since this algorithm does not explicitly account for diversity, we summarized using several durations, to assess the required summary length for this type of summarization to achieve the same classification performance of full songs or generic summarization. We report these results in Table X   We can see that this type of summarization reaches the performance of generic summaries (30 seconds) and full songs when the summary duration reaches 80 seconds (89.2% accuracy). This means that, for a human-oriented summary to be as descriptive and discriminative as a generic summary, an additional 50 seconds (2.67 times the length of the original) are needed. Even though the starting point of this contiguous summary is carefully selected by this algorithm, it still lacks diversity because of its contiguous nature, hindering classification accuracy for this summarizer. Naturally, by extending summary duration, summaries include more diverse information, eventually achieving the accuracy of full songs.

K. Structurally segmented sentences
Another form of human-oriented summarization is achieved by using generic summarization operating on structurally segmented sentences, done according to what humans might consider to be structurally meaningful segments. After structural segmentation, we fed each of the 5 generic algorithms with the resulting sentences instead of fixed-size ones and truncated the summary at 30 seconds, when necessary. The parameterization used for these experiments was the one that yielded the best results in the previous experiments for each algorithm.
The accuracy results for GRASSHOPPER, LexRank, LSA, MMR, and Support Sets were, respectively, 82.64%, 83.76%, 81.84%, 82.40%, and 83.84%. Even though structurally segmented sentences slightly improve performance, when considering classification accuracy, they are still outperformed by fixed-size segmentation. The best algorithm can only achieve 83.84% accuracy. This is because these sentences are much longer, therefore harming diversity in summaries. Furthermore, important content in structural sentences can always be extracted when using smaller fixed-size sentences. Thus, using smaller sentences, prevents the selection of redundant content.

VII. DISCUSSION
We ran the Wilcoxon signed-ranked test on all of the confusion matrices presented above against the full songs scenario. The continuous sections' p-values were 3.104×10 −4 , 3.628 × 10 −3 , and 2.858 × 10 −5 for the 30-second beginning, middle, and end sections of the songs, respectively, which means that they differ markedly from using full songs (as can also be seen by the accuracy drops they cause). The summaries, however, were very close to full songs, in terms of accuracy. The p-values for GRASSHOPPER, LexRank, LSA, MMR, and Support Sets were 0.10, 0.09, 0.16, 0.20, and 0.22, respectively. Thus, statistically speaking, using any of these 30-second summaries does not significantly differ from using full songs for classification (considering 95% confidence intervals). Furthermore, the p-values for 20-second LSA summaries and for 10-second Support Sets summaries were 0.06 and 0.08, respectively, with the remaining p-values of increasing summary sizes also being superior to 0.05. Thus, statistically speaking, generic summarization (in some cases) does not significantly differ from using full songs for classification, for summaries as short as 10 seconds (considering a 95% confidence interval). This is noteworthy, considering that the average song duration in this dataset is 283 seconds, which means that we achieve similar levels of classification performance using around 3.5% of the data. Human-oriented summarization is able to achieve these performance levels, but only at 50-second summaries and with a p-value of 0.055, barely over the 0.05 threshold. However, the 60-second summaries produced by this algorithm cannot reach that threshold. Only at 80 seconds is a comfortable p-value (0.38) for the 95% confidence interval attained.
Although every algorithm creates summaries in a different way, they all tend to include relevant and diverse sentences. This compensates their reduced lengths (up to 30 seconds of audio) allowing those clips to be representative of the whole musical pieces, from an automatic consumption view, as demonstrated by our experiments. Moreover, choosing the best 30-second contiguous segments is highly dependent on the genres in the dataset and tasks it will be used for, which is another reason for preferring summaries over those segments. The more varied the dataset, the less likely a fixed continuous section extraction method is to produce representative enough clips. Bass and Trance were the most influenced genres, by summarization, in these experiments. These are styles with very well defined structural borders, and a very descriptive structural element -the drop. The lack of that same element in a segment markedly hinders classification performance, suggesting that any genre with similar characteristics may also benefit from this type of summarization. It is also worth restating that Hip hop and Indie Rock were very positively influenced by summarization, regarding classification performance improvements over using full songs. This shows that, sometimes, classification on summarized music can even outperform using the whole data from the original signal. We also demonstrated that generic summarization using fixed-size sentences, that is, summarization not specifically oriented towards human consumption greatly outperforms human-oriented summarization approaches for the classification task.
Summarizing music prior to the classification task also takes time, but we do not claim it is worth doing it every time we are about do perform a MIR task. The idea is to compute summarized datasets offline for future use in any task that can benefit from them (e.g., music classification). Currently, sharing music datasets for MIR research purposes is very limited in many aspects, due to copyright issues. Usually, datasets are shared through features extracted from (30-second) continuous clips. That practice has drawbacks, such as: those 30 seconds may not contain the most relevant information and may even be highly redundant; and the features provided may not be the ones a researcher needs for his/her experiments. Summarizing datasets this way also helps avoiding copyright issues (because summaries are not created in a way enjoyable by humans) and still provide researchers with the most descriptive parts (according to each summarizer) of the signal itself, so that many different kinds of features can be extracted from them.

VIII. CONCLUSIONS AND FUTURE WORK
We showed that generic summarization algorithms perform well when summarizing music datasets about to be classified. The resulting summaries are remarkably more descriptive of the whole songs than their continuous segments (of the same duration) counterparts. Sometimes, these summaries are even more discriminative than the full songs. We also presented an argument stating some advantages in sharing summarized datasets within the MIR community.
An interesting research direction would be to automatically determine the best vocabulary size for each song. Testing summarization's performance on different classification tasks (e.g., with more classes) is also necessary to further strengthen our conclusions. More comparisons with non-contiguous humanoriented summaries should also be done. More experimenting should be done in other MIR tasks that also make use of only a portion of the whole signal.