Compositional chroma estimation using powered Euclidean distance

We develop a theoretical framework for representation and automated generation of South Indian classical music. The foundational part of the latter is based on symbolic dynamics and is implemented by translating the lexicographic rules for a rāga to constraints on a Markov chain whose state space is a layered graph. We analyze the statistical properties of this Markov chain from the point of view of information theory. We also develop several tools in music signal processing, such as, (a) a procedure for automated generation of gamakas or ornamental notes, unique to South Indian classical music, (b) rhythm synchronization, and (c) an algorithm for perceptual scale shifts. The Online Supplement has computer synthesized music from Śankarābharaṇa rāga, rhythm synchronized music for a Mōhana rāga composition, original and synthesized gamakas over madhyama for Begaḍa and Nīlāmbari rāgas, and perceptually scale shifted rāgas Hindoḷa, Madhyamāvati, Śuddha sāveri, and Udaya ravi chandrikā rāgas from the base Mōhana rāga.

Non-negative Matrix Factorisation (NMF) is a popular tool in musical signal processing. However, problems using this methodology in the context of Automatic Music Transcription (AMT) have been noted resulting in the proposal of supervised and constrained variants of NMF for this purpose. Group sparsity has previously been seen to be effective for AMT when used with stepwise methods. In this paper group sparsity is introduced to supervised NMF decompositions and a dictionary tuning approach to AMT is proposed based upon group sparse NMF using the β-divergence. Experimental results are given showing improved AMT results over the state-of-the-art NMF-based AMT system.

Learning a high-dimensional dense representation for vocabulary terms, also known as a word embedding, has recently attracted much attention in natural language processing and information retrieval tasks. The embedding vectors are typically learned based on term proximity in a large corpus. This means that the objective in well-known word embedding algorithms, e.g., word2vec, is to accurately predict adjacent word(s) for a given word or context. However, this objective is not necessarily equivalent to the goal of many information retrieval (IR) tasks. The primary objective in various IR tasks is to capture relevance instead of term proximity, syntactic, or even semantic similarity. This is the motivation for developing unsupervised relevance-based word embedding models that learn word representations based on query-document relevance information. In this paper, we propose two learning models with different objective functions; one learns a relevance distribution over the vocabulary set for each query, and the other classifies each term as belonging to the relevant or non-relevant class for each query. To train our models, we used over six million unique queries and the top ranked documents retrieved in response to each query, which are assumed to be relevant to the query. We extrinsically evaluate our learned word representation models using two IR tasks: query expansion and query classification. Both query expansion experiments on four TREC collections and query classification experiments on the KDD Cup 2005 dataset suggest that the relevance-based word embedding models significantly outperform state-of-the-art proximity-based embedding models, such as word2vec and GloVe.

Deep Learning has emerged as a powerful set of computational methods achieving great results in a variety of different tasks. Music signal processing, a field with rich commercial applications, seems to benefit too from this data-driven approach. In this paper a review of the state of the art Deep Learning methods applied on music signal processing is provided. A special focus is given in music information retrieval and music generation. In addition, possible future research directions are discussed.

Representing data residing on a graph as a linear combination of building block signals can enable efficient and insightful visual or statistical analysis of the data, and such representations prove useful as regularizers in signal processing and machine learning tasks. Designing collections of building block signals -- or more formally, dictionaries of atoms -- that specifically account for the underlying graph structure as well as any available representative training signals has been an active area of research over the last decade. In this article, we survey a particular class of dictionaries called localized spectral graph filter frames, whose atoms are created by localizing spectral patterns to different regions of the graph. After showing how this class encompasses a variety of approaches from spectral graph wavelets to graph filter banks, we focus on the two main questions of how to design the spectral filters and how to select the center vertices to which the patterns are localized. Throughout, we emphasize computationally efficient methods that ensure the resulting transforms and their inverses can be applied to data residing on large, sparse graphs. We demonstrate how this class of transform methods can be used in signal processing tasks such as denoising and non-linear approximation, and provide code for readers to experiment with these methods in new application domains.

The authors derive efficient parallel algorithms on a VLSI array for several signal and image processing tasks. The parallel architecture has n/sup 2/ memory modules storing the input data which is accessed by n PEs. They consider several signal processing tasks such as k selection, median filtering, labeling 0/1 image, etc. For these problems, linear speedup is obtained, compared to a single-processor system. For the k selection problem they also provide an efficient parallel solution using log transformation on the data. The expected time of this method is O(n) with a small constant factor. The proposed array is suitable for general-purpose signal processing and can be implemented in VLSI using a limited chip set.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>

The following report presents some of the ongoing projects that are taking place in the group’s laboratory. One of the noteable characteristics of this group is the extensive research spectrum, the plurality of research areas that are being studied by it’s members, such as Music Information Retrieval, Signal Processing and New Interfaces for Musical Expression.

This paper addresses feature extraction for automatic chord recognition systems. Most chord recognition systems use chroma features as a front-end and some kind of classifier (HMM, SVM or template matching). The vast majority of feature extraction approaches are based on mapping frequency bins from spectrum or constant-Q spectrum to chroma bins. In this work a set of new chroma features that are based on the time-frequency reassignment (TFR) technique is investigated. The proposed feature set was evaluated on the commonly used Beatles dataset and proved to be efficient for the chord recognition task, outperforming standard chroma.

Content representation is a fundamental task in information retrieval. Representation learning is aimed at capturing features of an information object in a low-dimensional space. Most research on representation learning for heterogeneous information networks (HINs) focuses on static HINs. In practice, however, networks are dynamic and subject to constant change. In this article, we propose a novel and scalable representation learning model, M-DHIN , to explore the evolution of a dynamic HIN. We regard a dynamic HIN as a series of snapshots with different time stamps. We first use a static embedding method to learn the initial embeddings of a dynamic HIN at the first time stamp. We describe the features of the initial HIN via metagraphs, which retains more structural and semantic information than traditional path-oriented static models. We also adopt a complex embedding scheme to better distinguish between symmetric and asymmetric metagraphs. Unlike traditional models that process an entire network at each time stamp, we build a so-called change dataset that only includes nodes involved in a triadic closure or opening process, as well as newly added or deleted nodes. Then, we utilize the above metagraph-based mechanism to train on the change dataset. As a result of this setup, M-DHIN is scalable to large dynamic HINs since it only needs to model the entire HIN once while only the changed parts need to be processed over time. Existing dynamic embedding models only express the existing snapshots and cannot predict the future network structure. To equip M-DHIN with this ability, we introduce an LSTM-based deep autoencoder model that processes the evolution of the graph via an LSTM encoder and outputs the predicted graph. Finally, we evaluate the proposed model, M-DHIN , on real-life datasets and demonstrate that it significantly and consistently outperforms state-of-the-art models.

Stacked auto-encoders based visual features for speech/music classification

In the last few years various musicological studies on the Byzantine Chant (BC) have been made via computer analysis. Due to the BC vocal nature and the pitch variability of the BC notes during the performance, the value of a music interval between two such notes in a certain BC mode - “echos”, is ambiguous. Because of that pitch variability a new method of BC intervals measurement is proposed, based on statistical processing of the pitch track that can be done after a proper pitch analysis of the BC recordings. The music scale of a particular BC echos is computed from the histogram of the pitch values by applying an algorithm that calculates a Gaussian Mixture Model approximating the music scale of that BC echos. This new method is applied, as an example, to ten BC recordings - performances of the Cherubic Hymn composed in the fourth echos (aghia) by Petros Lampadarios. The recordings come from the DAMASKINOS Corpus of University of Athens and the analysis results are in agreement with both the theoretical values of the Patriarchal Committee 1881 (approximation of up to 1 moria) and the findings in other similar studies. This method of analysis can be a base for both musicological analysis and music signal processing.

Detecting multiple pitches (F0s) and segregating musical instrument lines from monaural recordings of contrapuntal polyphonic music into separate tracks is a difficult problem in music signal processing. Applications include audio-to-MIDI conversion, automatic music transcription, and audio enhancement and transformation. Past attempts at separation have been limited to separating two harmonic signals in a contrapuntal duet (Maher, 1990) or several harmonic signals in a single chord (Virtanen and Klapuri, 2001, 2002). Several researchers have attempted polyphonic pitch detection (Klapuri, 2001; Eggink and Brown, 2004a), predominant melody extraction (Goto, 2001; Marolt, 2004; Eggink and Brown, 2004b), and instrument recognition (Eggink and Brown, 2003). Our solution assumes that each instrument is represented as a time-varying harmonic series and that errors can be corrected using prior knowledge of instrument spectra. Fundamental frequencies (F0s) for each time frame are estimated from input spectral data using an Expectation-Maximization (EM) based algorithm with Gaussian distributions used to represent the harmonic series. Collisions (i.e., overlaps) between instrument harmonics, which frequently occur, are predicted from the estimated F0s. The uncollided harmonics are matched to ones contained in a pre-stored spectrum library in order that each F0‘s harmonic series is assigned to the appropriate instrument. Corrupted harmonics are restored using data taken from the library. Finally, each voice is additively resynthesized to a separate track. This algorithm is demonstrated for a monaural signal containing three contrapuntal musical instrument voices with distinct timbres.KeywordsHarmonic SpectrumHarmonic SeriesMusic Information RetrievalPitch DetectionSeparate TrackThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

The traditional computer-aided design (CAD) software is often not realistic in rendering scenes, such as light, materials, shadows, and other details, which leads designers to spend a lot of energy on manual adjustment.The rendering algorithm based on deep learning (DL) can automatically learn the complex output, and the effect is usually better than traditional methods.In this article, a convolutional neural network (GA-CNN) based on genetic algorithm (GA) optimization is proposed, and it is combined with CAD data and applied to the rendering of environmental art design scenes.This method uses GA to optimize the parameters of CNN to improve its learning effect and generalization ability.From the overall trend, the response speed of GA-CNN algorithm and support vector machine (SVM).The response speed of the traditional CNN method is relatively slow, and the fluctuation range is large.This is because traditional CNN needs a lot of convolution operations and feature extraction when dealing with artistic scenes in complex environments, which leads to a long calculation time.In practical application, it is necessary to select appropriate algorithms and parameter settings according to specific task requirements to achieve the best results.

Music technologies will open the future up to new ways of enjoying music both in terms of music appreciation and music creation. In this keynote speech, I introduce the frontiers of music technologies by showing some practical examples to demonstrate how end users can benefit from music signal processing, music understanding technologies, singing synthesis technologies, and music interfaces. For example, a web service for active music listening, "Songle" (http://songle.jp), has analyzed more than 830,000 songs on music- or video-sharing services and facilitates deeper understanding of music and music-synchronized control of robot dancers. A web service for large-scale music browsing, "Songrium" (http://songrium.jp), allows users to explore music while seeing and utilizing various relations among more than 680,000 music video clips on video-sharing services. Singing synthesis technologies open up new possibilities in music creation. I conclude by discussing grand challenges.

Music signal processing is one of the active research area now a days. Identifying the musical instrument form the solo recordings is one of the applications under signal processing. Different types of algorithms have been proposed till now to identify the instrument. For this objective first of all we have to extract the features from sound samples and then we can use them for instrument identification. In this paper we are presenting one of efficient method to extract the features from sound sample that is Mel Frequency Cepsrtal Coefficient (MFCC). These MFCCs are closely related to behavior of human auditory system. While calculating these MFCCs we actually deal with spectral envelop. And hence it gives very distinctive features of the sound samples. So using these features for identification of musical instrument we can increase the accuracy of identification. In our work we have used five instruments and 30 samples of each instrument. In the training database total 90 samples (60% of total samples) and for testing database 60 (40%)samples are used. In the classification phase we have used K-Nearest Neighbor (K-NN) classifier. Proposed system gives the accuracy of 91.66% for Cello, Piano and Trumpet and accuracy of 83.33% for Flute and Violin.