Search

Previous research in music emotion recognition (MER) has tackled the inherent problem of subjectivity through the use of personalized models - models which predict the emotions that a particular user would perceive from music. Personalized models are trained in a supervised manner, and are tested exclusively with the annotations provided by a specific user. While past research has focused on model adaptation or reducing the amount of annotations required from a given user, we propose a methodology based on uncertainty sampling and query-by-committee, adopting prior knowledge from the agreement of human annotations as an oracle for active learning (AL). We assume that our disagreements define our personal opinions and should be considered for personalization. We use the DEAM dataset, the current benchmark dataset for MER, to pre-train our models. We then use the AMG1608 dataset, the largest MER dataset containing multiple annotations per musical excerpt, to re-train diverse machine learning models using AL and evaluate personalization. Our results suggest that our methodology can be beneficial to produce personalized classification models that exhibit different results depending on the algorithms' complexity. [ABSTRACT FROM AUTHOR]

Version identification systems aim to detect different renditions of the same underlying musical composition (loosely called cover songs). By learning to encode entire recordings into plain vector embeddings, recent systems have made significant progress in bridging the gap between accuracy and scalability, which has been a key challenge for nearly two decades. In this work, we propose to further narrow this gap by employing a set of data distillation techniques that reduce the embedding dimensionality of a pre-trained state-of-the-art model. We compare a wide range of techniques and propose new ones, from classical dimensionality reduction to more sophisticated distillation schemes. With those, we obtain 99% smaller embeddings that, moreover, yield up to a 3% accuracy increase. Such small embeddings can have an important impact in retrieval time, up to the point of making a real-world system practical on a standalone laptop. [ABSTRACT FROM AUTHOR]

Tagging a musical excerpt with an emotion label may result in a vague and ambivalent exercise. This subjectivity entangles several high-level music description tasks when the computational models built to address them produce predictions on the basis of a "ground truth". In this study, we investigate the relationship between emotions perceived in pop and rock music (mainly in Euro-American styles) and personal characteristics from the listener, using language as a key feature. Our goal is to understand the influence of lyrics comprehension on music emotion perception and use this knowledge to improve Music Emotion Recognition (MER) models. We systematically analyze over 30K annotations of 22 musical fragments to assess the impact of individual differences on agreement, as defined by Krippendorff's α coefficient. We employ personal characteristics to form group-based annotations by assembling ratings with respect to listeners' familiarity, preference, lyrics comprehension, and music sophistication. Finally, we study our group-based annotations in a two-fold approach: (1) assessing the similarity within annotations using manifold learning algorithms and unsupervised clustering, and (2) analyzing their performance by training classification models with diverse "ground truths". Our results suggest that a) applying a broader categorization of taxonomies and b) using multi-label, group-based annotations based on language, can be beneficial for MER models. [ABSTRACT FROM AUTHOR]

Unison singing is the name given to an ensemble of singers simultaneously singing the same melody and lyrics. While each individual singer in a unison sings the same principle melody, there are slight timing and pitch deviations between the singers, which, along with the ensemble of timbres, give the listener a perceived sense of "unison". In this paper, we present a study of unison singing in the context of choirs; utilising some recently proposed deeplearning based methodologies, we analyse the fundamental frequency (F0) distribution of the individual singers in recordings of unison mixtures. Based on the analysis, we propose a system for synthesising a unison signal from an a cappella input and a single voice prototype representative of a unison mixture. We use subjective listening tests to evaluate perceptual factors of our proposed system for synthesis, including quality, adherence to the melody as well the degree of perceived unison. [ABSTRACT FROM AUTHOR]

This paper addresses the extraction of multiple F0 values from polyphonic and a cappella vocal performances using convolutional neural networks (CNNs). We address the major challenges of ensemble singing, i.e., all melodic sources are vocals and singers sing in harmony. We build upon an existing architecture to produce a pitch salience function of the input signal, where the harmonic constant-Q transform (HCQT) and its associated phase differentials are used as an input representation. The pitch salience function is subsequently thresholded to obtain a multiple F0 estimation output. For training, we build a dataset that comprises several multi-track datasets of vocal quartets with F0 annotations. This work proposes and evaluates a set of CNNs for this task in diverse scenarios and data configurations, including recordings with additional reverb. Our models outperform a state-of-the-art method intended for the same music genre when evaluated with an increased F0 resolution, as well as a general-purpose method for multi-F0 estimation. We conclude with a discussion on future research directions. [ABSTRACT FROM AUTHOR]

Generating continuous f0 annotations for tasks such as melody extraction and multiple f0 estimation typically involves running a monophonic pitch tracker on each track of a multitrack recording and manually correcting any estimation errors. This process is labor intensive and time consuming, and consequently existing annotated datasets are very limited in size. In this paper we propose a framework for automatically generating continuous f0 annotations without requiring manual refinement: the estimate of a pitch tracker is used to drive an analysis/synthesis pipeline which produces a synthesized version of the track. Any estimation errors are now reflected in the synthesized audio, meaning the tracker's output represents an accurate annotation. Analysis is performed using a wide-band harmonic sinusoidal modeling algorithm which estimates the frequency, amplitude and phase of every harmonic, meaning the synthesized track closely resembles the original in terms of timbre and dynamics. Finally the synthesized track is automatically mixed back into the multitrack. The framework can be used to annotate multitrack datasets for training learning-based algorithms. Furthermore, we show that algorithms evaluated on the automatically generated/ annotated mixes produce results that are statistically indistinguishable from those they produce on the original, manually annotated, mixes. We release a software library implementing the proposed framework, along with new datasets for melody, bass and multiple f0 estimation. [ABSTRACT FROM AUTHOR]

Score information has been shown to improve music source separation when included into non-negative matrix factorization (NMF) frameworks. Recently, deep learning approaches have outperformed NMF methods in terms of separation quality and processing time, and there is scope to extend them with score information. In this paper, we propose a score-informed separation system for classical music that is based on deep learning. We propose a method to derive training features from audio files and the corresponding coarsely aligned scores for a set of classical music pieces. Additionally, we introduce a convolutional neural network architecture (CNN) with the goal of estimating time-frequency masks for source separation. Our system is trained with synthetic renditions derived from the original scores and can be used to separate real-life performances based on the same scores, provided a coarse audio-to-score alignment. The proposed system achieves better performance (SDR and SIR) and is less computationally intensive than a score-informed NMF system on a dataset comprising Bach chorales. [ABSTRACT FROM AUTHOR]

We present a study, carried out on 241 participants, which investigates on classical music material the agreement of listeners on perceptual music aspects (related to emotion, tempo, complexity, and instrumentation) and the relationship between listener characteristics and these aspects. For the currently popular task of music emotion recognition, the former question is particularly important when defining a ground truth of emotions perceived in a given music collection. We characterize listeners via a range of factors, including demographics, musical inclination, experience, and education, and personality traits. Participants rate the music material under investigation, i.e., 15 expert-defined segments of Beethoven's 3rd symphony, "Eroica", in terms of 10 emotions, perceived tempo, complexity, and number of instrument groups. Our study indicates only slight agreement on most perceptual aspects, but significant correlations between several listener characteristics and perceptual qualities. [ABSTRACT FROM AUTHOR]