Beyond Error Tolerance: Finding Thematic Similarities in Music Digital Libraries

Metadata research for music digital libraries has traditionally focused on genre. Despite its potential for improving the ability of users to better search and browse music collections, music subject metadata is an unexplored area. The objective of this study is to expand the scope of music metadata research, in particular, by exploring music subject classification based on user interpretations of music. Furthermore, we compare this previously unexplored form of user data to lyrics at subject prediction tasks. In our experiment, we use datasets consisting of 900 songs annotated with user interpretations. To determine the significance of performance differences between the two sources, we applied Friedman's ANOVA test on the classification accuracies. The results show that user-generated interpretations are significantly more useful than lyrics as classification features (p < 0.05). The findings support the possibility of exploiting various existing sources for subject metadata enrichment in music digital libraries.

We continue our work on the automatic mining of user-created music reviews towards the goal of connecting user opinions to music objects in Music Digital Libraries (MDL). We demonstrate an experimental system which automatically discovered the key descriptive patterns that differentiated positive from negative reviews which helps us to better understand our successful Phase I results. Comparison to an earlier study indicates an important consistency across projects that warrants further investigation.

This paper introduces the Raga Based Music Classification as a part of Music Information Retrieval (MIR) systems for retrieving musical information. The music has been more than just a sound that entertains us. MIR users like musicology researchers, musicians, music learners, music therapists, etc., require not just the sound but the actual strength of the music that controls our body, motivates us, cures mental illness, etc. The backbone of any form of music is its raga. A raga is a particular combination of notes with certain laws which, when carefully observed, preserves and safeguards their integrity and create wonders in music. The existing MIR systems and audio search engines concentrate on content based audio mining and not ragas. This paper discusses the classification of ragas and how it differs in Carnatic, Hindustani and Western musical forms. It also describes the various features that can be extracted from an audio data to represent a raga. Finally, we propose a system architecture to perform raga based classification of music.

Music Recognition System has becoming popular these days. They are based on either audio or symbolic method, which compares the user's query with the existing music database. The investigation has shown that the audio-based method system is good for storing sound waves. However, the limitation is to illustrate the content of the music. The symbolic-based method system doing well in representing the content of the music, including recognize similar patterns, but limitation in creating the music, e.g., Electronic Music. We also carried some detailed tests based on the previous Neural Network systems, with Music Definition Language (MDL) and Music Manipulation Language (MML). Furthermore, we have tested our previous classification system with new melody query, to see how it can handle with external music pieces, which beyond the self-testing from a self-organising-map. The conclusion is that our system can classify variation type, including key variations, expansion and reduction, which is better than those existing Music Information Retrieval (MIR) systems.

The design and implementation of Harbin Institute of Technology--Digital Music Library (HIT-DML) is presented in this paper. Firstly, a novel framework, a music data model, and a query language are proposed as the theoretical foundation of the library. Secondly, music computing algorithms used in the library for feature extracting and matching are described. In addition, indices are introduced for both mining themes of music objects and accelerating content-based information retrieval. Finally, experimental results on the indices and the current development of the library are provided. HIT-DML is distinguished by the following points. First, it is inherently based on database systems, and combines database technologies with multimedia technologies seamlessly. Musical data are structurally stored. Second, it has a solid theoretical foundation, from framework and data model to query language. Last, it can retrieve musical information based on content against different kinds of musical instruments. The indices used, also power the library.

This study was conducted at the Music Library of the Ministry of Information (ML-MOI) and the Music Library of the Cultural Affairs (ML-CA) of Eritrea. The main aim of the study was to designan appropriate Indigenous Music Information Storage and Retrieval System for Eritrea. A quantitative approach was mainly used to obtain data from a purposefully selected sample. The qualitative approach was also used in some research stages. Methods used included documentanalysis, questionnaire, personal interview, focus group discussion and observation. It was found that music data was recorded in a register, which also served as a finding aid. Users’ preference was to search for music by artist, title, producer, occasion, subject or language.Unfortunately this was not possible with the system that was being used since it only allowed searching through the name or initials of an artist. The paper recommends the establishment of an automated indigenous storage and retrieval system to enable registration of indigenous knowledge information, offering facilities for indexing, searching and report generation.

This paper describes work in progress on an ethnographic study of the searching/browsing techniques employed by people in the researcher’s local community. The insights provided by this type of study can inform the development of searching/browsing support for music digital libraries.

This poster describes a real-time audio-based automatic music genre classifier for use in organizing, browsing, and searching musical digital libraries. A decision tree classifier trained on a 40-dimension feature space is used to categorize music into one of 14 different genres with the results being displayed to a continuously updating user interface.

Recently, several music information retrieval (MIR) systems which retrieve musical pieces by the user's singing voice have been developed. All of these systems use only melody information for retrieval, although lyrics information is also useful for retrieval. In this paper, we propose a new MIR system that uses both lyrics and melody information. First, we propose a new lyrics recognition method. A finite state automaton (FSA) is used as recognition grammar, and about 86% retrieval accuracy was obtained. We also develop an algorithm for verifying a hypothesis output by a lyrics recognizer. Melody information is extracted from an input song using several pieces of information of the hypothesis, and a total score is calculated from the recognition score and the verification score. From the experimental results, 95.0% retrieval accuracy was obtained with a query consisting of five words.

Music information retrieval (MIR) systems tend to fall into two camps: that camp developing cataloging and providing advanced access systems for large collections of music and that camp developing specific query or access mechanisms. We have started to merge these camps by integrating Variations2, which provides access to a digitized portion of Indiana University's vast music library, with Michigan's VocalSearch, which provides a query-by-humming (QBH) search engine. The joint system, V2V, demonstrates how QBH can be used in connection with a large number of holdings in a real-world environment.

In this paper, we present an organized survey of the existing literature on music information retrieval systems in which descriptor features are extracted directly from the compressed audio files, without prior decompression to pulse-code modulation format. Avoiding the decompression step and utilizing the readily available compressed-domain information can significantly lighten the computational cost of a music information retrieval system, allowing application to large-scale music databases. We identify a number of systems relying on compressed-domain information and form a systematic classification of the features they extract, the retrieval tasks they tackle and the degree in which they achieve an actual increase in the overall speed—as well as any resulting loss in accuracy. Finally, we discuss recent developments in the field, and the potential research directions they open toward ultra-fast, scalable systems.

Perceived predominant melody of music is the most convenient and memorable description and can be used for content-based music retrieval. However, including human voice and multiple musical instruments playing together, it is difficult to extract a predominant contour of pitches directly from MP3 music recordings. In order to build a similarity music melody model, several audio files, whose melody is perceptually similar, are collected in our experiment. Many features are extracted and a melodic similarity model is defined through analyzing each feature and combinations of them. The melody model is evaluated based on classification results of six categories of Chinese folk music using Support Vector Machine. The experiment results show that 36-dimensional Constant Q Transform (CQT) feature can represent the melody of audio music pieces accurately. Further more, the classification results for audio data with similar melody are good enough to be used in audio music classification or segmentation and subsequently are very helpful in music information retrieval system.

This paper deals with the Music/Speech classification problem, starting from a set of features extracted directly from compressed audio data. The proposed classification system is able to label audio sequences stored as compressed MPEG layer III files. Decoding and analyzing in a unique stage is a fundamental tool for audio streaming applications, such as real time classification. Moreover, the techniques described herein provide useful tools in the management (data tagging, summarization, etc.) of a digital music library. The adopted set of short-time features are computed from the spectral information available in the decoding stage. In this paper, we show that for the classification problem at hand this set of features is redundant and can be dramatically pruned. To this aim we used an optimization strategy based on principal component analysis and genetic algorithms. The results show a very interesting classification accuracy using just one short-time feature.

MiDiLiB is a six year research project on digital music libraries funded by the German Research Foundation (DFG) as a part of the Distributed Processing and Delivery of Digital Documents (V3D2) research initiative. MiDiLiB’s main focus is the development of content-based retrieval algorithms for both score- and waveform-based music. In this paper we give an overview of our research results, describe several prototypical systems for content-based music retrieval which have been developed during the project, and discuss applications of the presented techniques in the context of today’s and future digital music libraries.

Musicologists and musicians often would like to search by keys in a digital music library. In this paper, we introduce a new key-finding algorithm that can be applied to music in both symbolic and audio formats. The algorithm, which is based on a Hidden Markov Model (HMM), provides two stages of key-finding output; the first one referring to local keys and the second one to the global key. We describe the input, the two output stages, and the parameters of the model. In particular, we describe two configurable parameters, the transition probability distributions, which are based on a matrix of neighbouring keys, and the emission probability distributions, which make use of established key profiles. We discuss the local key-finding capabilities of the algorithm, presenting an example analysis of the Prelude Op. 28 No. 20 in C minor by Chopin, showing the local key regions obtained using different key profiles. We evaluate the global key-finding capabilities of the model, using an existing dataset and six well-known key profiles as different model parameters. Since different key profiles will tend to err or misclassify in different ways and across different pieces, we train an ensemble method with the predictions from all the key profiles (6) through our model. We show that the ensemble method achieves state-of-the-art performance for major and overall keys, however, it still underperforms the state-of-the-art for minor keys.