Search

Music Information Retrieval (MIR) has seen a recent surge in deep learning-based approaches, which often involve encoding symbolic music (i.e., music represented in terms of discrete note events) in an image-like or languagelike fashion. However, symbolic music is neither an image nor a sentence, and research in the symbolic domain lacks a comprehensive overview of the different available representations. In this paper, we investigate matrix (piano roll), sequence, and graph representations and their corresponding neural architectures, in combination with symbolic scores and performances on three piece-level classification tasks. We also introduce a novel graph representation for symbolic performances and explore the capability of graph representations in global classification tasks. Our systematic evaluation shows advantages and limitations of each input representation. Our results suggest that the graph representation, as the newest and least explored among the three approaches, exhibits promising performance, while being more light-weight in training. [ABSTRACT FROM AUTHOR]

In jazz, measuring harmonic similarity is complicated by the common practice of reharmonization -- the altering or substitution of chords without fundamentally changing the piece's harmonic identity. This is analogous to natural language processing tasks where synonymous terms can be used interchangeably without significantly modifying the meaning of a text. Our approach to modeling harmonic similarity borrows from NLP techniques, such as distributional semantics, by embedding chords into a vector space using a co-occurrence matrix. We show that the method can robustly detect harmonic similarity between songs, even when reharmonized. The co-occurrence matrix is computed from a corpus of symbolic jazz-chord progressions, and the result is a map from chords into vectors. A song's harmony can then be represented as a piecewiselinear path constructed from the cumulative sum of its chord vectors. For any two songs, their harmonic similarity can be measured as the minimal surface membrane area between their vector paths. Using a dataset of jazz contrafacts, we show that our approach reduces the median rank of matches from 318 to 18 compared to a baseline approach using pitch class vectors. [ABSTRACT FROM AUTHOR]

We present FiloBass: a novel corpus of music scores and annotations which focuses on the important but often overlooked role of the double bass in jazz accompaniment. Inspired by recent work that sheds light on the role of the soloist, we offer a collection of 48 manually verified transcriptions of professional jazz bassists, comprising over 50,000 note events, which are based on the backing tracks used in the FiloSax dataset. For each recording we provide audio stems, scores, performance-aligned MIDI and associated metadata for beats, downbeats, chord symbols and markers for musical form. We then use FiloBass to enrich our understanding of jazz bass lines, by conducting a corpus-based musical analysis with a contrastive study of existing instructional methods. Together with the original FiloSax dataset, our work represents a significant step toward a fully annotated performance dataset for a jazz quartet setting. By illuminating the critical role of the bass in jazz, this work contributes to a more nuanced and comprehensive understanding of the genre. [ABSTRACT FROM AUTHOR]

Computational models of expressive piano performance rely on attributes like tempo, timing, dynamics and pedalling. Despite some promising models for performance assessment and performance rendering, results are limited by the scale, breadth and uniformity of existing datasets. In this paper, we present ATEPP, a dataset that contains 1000 hours of performances of standard piano repertoire by 49 world-renowned pianists, organized and aligned by compositions and movements for comparative studies. Scores in MusicXML format are also available for around half of the tracks. We first evaluate and verify the use of transcribed MIDI for representing expressive performance with a listening evaluation that involves recent transcription models. Then, the process of sourcing and curating the dataset is outlined, including composition entity resolution and a pipeline for audio matching and solo filtering. Finally, we conduct baseline experiments for performer identification and performance rendering on our dataset, demonstrating its potential in generalizing expressive features of individual performing style. [ABSTRACT FROM AUTHOR]

This paper makes several contributions to automatic lyrics transcription (ALT) research. Our main contribution is a novel variant of the Multistreaming Time-Delay Neural Network (MTDNN) architecture, called MSTRE-Net, which processes the temporal information using multiple streams in parallel with varying resolutions keeping the network more compact, and thus with a faster inference and an improved recognition rate than having identical TDNN streams. In addition, two novel preprocessing steps prior to training the acoustic model are proposed. First, we suggest using recordings from both monophonic and polyphonic domains during training the acoustic model. Second, we tag monophonic and polyphonic recordings with distinct labels for discriminating non-vocal silence and music instances during alignment. Moreover, we present a new test set with a considerably larger size and a higher musical variability compared to the existing datasets used in ALT literature, while maintaining the gender balance of the singers. Our best performing model sets the state-of-the-art in lyrics transcription by a large margin. For reproducibility, we publicly share the identifiers to retrieve the data used in this paper. [ABSTRACT FROM AUTHOR]

This paper proposes a deep convolutional neural network for performing note-level instrument assignment. Given a polyphonic multi-instrumental music signal along with its ground truth or predicted notes, the objective is to assign an instrumental source for each note. This problem is addressed as a pitch-informed classification task where each note is analysed individually. We also propose to utilise several kernel shapes in the convolutional layers in order to facilitate learning of timbre-discriminative feature maps. Experiments on the MusicNet dataset using 7 instrument classes show that our approach is able to achieve an average F-score of 0.904 when the original multi-pitch annotations are used as the pitch information for the system, and that it also excels if the note information is provided using third-party multi-pitch estimation algorithms. We also include ablation studies investigating the effects of the use of multiple kernel shapes and comparing different input representations for the audio and the note-related information. [ABSTRACT FROM AUTHOR]

The Filosax dataset is a large collection of specially commissioned recordings of jazz saxophonists playing with commercially available backing tracks. Five participants each recorded themselves playing the melody, interpreting a transcribed solo and improvising on 48 tracks, giving a total of around 24 hours of audio data. The solos are annotated both as individual note events with physical timing, and as sheet music with a metrical interpretation of the timing. In this paper, we outline the criteria used for choosing and sourcing the repertoire, the recording process and the semi-automatic transcription pipeline. We demonstrate the use of the dataset to analyse musical phenomena such as swing timing and dynamics of typical musical figures, as well as for training a source activity detection system and predicting expressive characteristics. Other potential applications include the modelling of jazz improvisation, performer identification, automatic music transcription, source separation and music generation. [ABSTRACT FROM AUTHOR]

We address the problem of confounding in the design of music classification experiments, that is, the inability to distinguish the effects of multiple potential influencing variables in the measurements. Confounding affects the validity of conclusions at many levels, and so must be properly accounted for. We propose a procedure for characterising effects of confounding in the results of music classification experiments by creating regulated test conditions through interventions in the experimental pipeline, including a novel resampling strategy. We demonstrate this procedure on the GTZAN genre collection, which is known to give rise to confounding effects. [ABSTRACT FROM AUTHOR]

The interpretability of a machine learning model is essential for gaining insight into model behaviour. While some machine learning models (e.g., decision trees) are transparent, the majority of models used today are still black-boxes. Recent work in machine learning aims to analyse these models by explaining the basis of their decisions. In this work, we extend one such technique, called local interpretable model-agnostic explanations, to music content analysis. We propose three versions of explanations: one version is based on temporal segmentation, and the other two are based on frequency and time-frequency segmentation. These explanations provide meaningful ways to understand the factors that influence the classification of specific input data. We apply our proposed methods to three singing voice detection systems: the first two are designed using decision tree and random forest classifiers, respectively; the third system is based on convolutional neural network. The explanations we generate provide insights into the model behaviour. We use these insights to demonstrate that despite achieving 71.4% classification accuracy, the decision tree model fails to generalise. We also demonstrate that the model-agnostic explanations for the neural network model agree in many cases with the model-dependent saliency maps. The experimental code and results are available online. [ABSTRACT FROM AUTHOR]

Though many past works have tried to cluster expressive timing within a phrase, there have been few attempts to cluster features of expressive timing with constant dimen- sions regardless of phrase lengths. For example, used as a way to represent expressive timing, tempo curves can be regressed by a polynomial function such that the number of regressed polynomial coefficients remains constant with a given order regardless of phrase lengths. In this paper, clustering the regressed polynomial coefficients is proposed for expressive timing analysis. A model selection test is presented to compare Gaussian Mixture Models (GMMs) fitting regressed polynomial coefficients and fitting expressive timing directly. As there are no expected results of clustering expressive timing, the proposed method is demonstrated by how well the expressive timing are approximated by the centroids of GMMs. The results show that GMMs fitting the regressed polynomial coefficients outperform GMMs fitting expressive timing directly. This conclusion suggests that it is possible to use regressed polynomial coefficients to represent expressive timing within a phrase and cluster expressive timing within phrases of different lengths. [ABSTRACT FROM AUTHOR]

Field recordings from ethnomusicological research since the beginning of the 20th century are available today in large digitised music archives. The application of music information retrieval and data mining technologies can aid large-scale data processing leading to a better understanding of the history of cultural exchange. In this paper we focus on folk and traditional music from the United Kingdom and study the correlation between spatial origins and musical characteristics. In particular, we investigate whether the geographical location of music recordings can be predicted solely from the content of the audio signal. We build a neural network that takes as input a feature vector capturing musical aspects of the audio signal and predicts the latitude and longitude of the origins of the music recording. We explore the performance of the model for different sets of features and compare the prediction accuracy between geographical regions of the UK. Our model predicts the geographical coordinates of music recordings with an average error of less than 120 km. The model can be used in a similar manner to identify the origins of recordings in large unlabelled music collections and reveal patterns of similarity in music from around the world. [ABSTRACT FROM AUTHOR]