Context-Aware Hidden Markov Models of Jazz Music
In this paper, a latent variable model is proposed, based on the Variable Markov Oracle that captures long-term temporal relationships between observations. A multi-level HMM-like model extracted from VMO, called VMO-HMM, is presented in this paper. The musical importance of the proposed model is demonstrated by modeling harmonic progressions of Jazz music. The proposed latent variable model is able to reveal functional harmony relations beyond chord labels and provides new theoretic insight into scale-chord relations and Jazz harmony improvisation as a model of music meta-creation
Re-visiting the Music Segmentation with Crowdsourcing
By Cheng-i Wang, Gautham J. Mysore, Shlomo Dubnov
Identifying boundaries in music structural segmentation is a well studied music information retrieval problem. The goal is to develop algorithms that automatically identify segmenting time points in music that closely matches human annotated data. The annotation itself is challenging due to its subjective nature, such as the degree of change that constitutes a boundary, the location of such boundaries, and whether a boundary should be assigned to a single time frame or a range of frames. Existing datasets have been annotated by small number of experts and the annotators tend to be constrained to specific definitions of segmentation boundaries. In this paper, we re-examine the annotation problem. We crowdsource the problem to a large number of annotators and present an analysis of the results. Our preliminary study suggests that although there is a correlation to existing datasets, this form of annotations reveals additional information such as stronger vs. weaker boundaries, gradual vs. sudden boundaries, and the difference in perception of boundaries between musicians and non-musicians. The study suggests that it could be worth re-defining certain aspects of the boundary identification in music structural segmentation problem with a broader definition.
Identifying boundaries in music structural segmentation is a well studied music information retrieval problem. The goal is to develop algorithms that automatically identify segmenting time points in music that closely matches human annotated data. The annotation itself is challenging due to its subjective nature, such as the degree of change that constitutes a boundary, the location of such boundaries, and whether a boundary should be assigned to a single time frame or a range of frames. Existing datasets have been annotated by small number of experts and the annotators tend to be constrained to specific definitions of segmentation boundaries. In this paper, we re-examine the annotation problem. We crowdsource the problem to a large number of annotators and present an analysis of the results. Our preliminary study suggests that although there is a correlation to existing datasets, this form of annotations reveals additional information such as stronger vs. weaker boundaries, gradual vs. sudden boundaries, and the difference in perception of boundaries between musicians and non-musicians. The study suggests that it could be worth re-defining certain aspects of the boundary identification in music structural segmentation problem with a broader definition.
Using music notation for teaching computer programming
By Eunjeong Stella Ko, K Lee
Eunjeong's research objectives are:
(1) to implement sequential models of data that can learn probabilistic rules of musical structure,
(2) to understand musical surface using neural networks/deep learning techniques, and
(3) to make people easily interact with the high dimensional data with auditory feedback and acoustic model.
Her research focuses on advancing the state of the art in deep learning algorithm modeling, and thereby improving several applications in the area of polyphonic music sequence generation, transcription and machine improvisation. Modeling real-world sequences often involves capturing long-term dependencies between the high-dimensional objects that compose such sequences. She try to follow from a standard deep learning approach to generate novel generation methodology. The main idea is to exploit the power of neural networks to learn a probabilistic description of sequences of symbols, that in turn can be used as a prior to improve the accuracy of information retrieval.
Eunjeong's research objectives are:
(1) to implement sequential models of data that can learn probabilistic rules of musical structure,
(2) to understand musical surface using neural networks/deep learning techniques, and
(3) to make people easily interact with the high dimensional data with auditory feedback and acoustic model.
Her research focuses on advancing the state of the art in deep learning algorithm modeling, and thereby improving several applications in the area of polyphonic music sequence generation, transcription and machine improvisation. Modeling real-world sequences often involves capturing long-term dependencies between the high-dimensional objects that compose such sequences. She try to follow from a standard deep learning approach to generate novel generation methodology. The main idea is to exploit the power of neural networks to learn a probabilistic description of sequences of symbols, that in turn can be used as a prior to improve the accuracy of information retrieval.
Method and apparatus for recommending reply message
By Eunjeong Stella Ko, D Kim, HJ Lee, J Rho, V Zubariev, T Yang, H Jung, HJ Jung
A device for transmitting a reply message is provided. The device includes a communication unit configured to receive a question message from another device, a controller configured to determine a category of the question message, a display unit configured to display a user interface (UI) for selecting data to be included in a reply message to the question message, according to the category, and a user input unit configured to receive a user input of selecting data to be included in the reply message through the UI, wherein the communication unit transmits the reply message including the data to the other device.
A device for transmitting a reply message is provided. The device includes a communication unit configured to receive a question message from another device, a controller configured to determine a category of the question message, a display unit configured to display a user interface (UI) for selecting data to be included in a reply message to the question message, according to the category, and a user input unit configured to receive a user input of selecting data to be included in the reply message through the UI, wherein the communication unit transmits the reply message including the data to the other device.