Fabien Gouyon

A Review of Automatic Rhythm Description Systems

This research was supported by the EU project FP6-507142 (SIMAC) and the national project Y99-INF sponsored by the Austrian Federal Ministryu of Education, Science and Culture in the form of a Start Research Prize. The Austrian Research Institute for Artificial Intelligence acknowledges the support of the Austrian Federal Ministries of Education, Science and Culture and of Transport, Innovationa and Technology.

Natural language processing of lyrics

We report experiments on the use of standard natural language processing (NLP) tools for the analysis of music lyrics. A significant amount of music audio has lyrics. Lyrics encode an important part of the semantics of a song, therefore their analysis complements that of acoustic and cultural metadata and is fundamental for the development of complete music information retrieval systems. Moreover, a textual analysis of a song can generate ground truth data that can be used to validate results from purely acoustic methods. Preliminary results on language identification, structure extraction, categorization and similarity searches suggests that a lot of profit can be gained from the analysis of lyrics.

Automatic Classification of Drum Sounds: A Comparison of Feature Selection Methods and Classification Techniques

We present a comparative evaluation of automatic classification of a sound database containing more than six hundred drum sounds (kick, snare, hihat, toms and cymbals). A preliminary set of fifty descriptors has been refined with the help of different techniques and some final reduced sets including around twenty features have been selected as the most relevant. We have then tested different classification techniques (instance-based, statistical-based, and tree-based) using ten-fold cross-validation. Three levels of taxonomic classification have been tested: membranes versus plates (super-category level), kick vs. snare vs. hihat vs. toms vs. cymbals (basic level), and some basic classes (kick and snare) plus some sub-classes -i.e. ride, crash, open-hihat, closed hihat, high-tom, medium-tom, low-tom- (sub-category level). Very high hit-rates have been achieved (99%, 97%, and 90% respectively) with several of the tested techniques.

Automatic extraction of drum tracks from polyphonic music signals

We propose an approach for extracting automatically time indexes of occurrences of percussive sounds in an audio signal taken from the popular music repertoire. The scheme is able to detect percussive sounds unknown a priori in a selective fashion. It is based on an analysis by synthesis technique, whereby the sound searched for is gradually synthesized from the signal itself. The possibility to extract different types of percussive sounds and their occurrences in the audio signal makes it possible to build a drum track representing the essential rhythmic component of a music piece. We present a systematic evaluation of the performance of our algorithm on a database of popular music titles. The system performs well on most of the cases (over 75%). We analyze the reasons for failure on the remaining cases, and propose solutions for yet improving the algorithm. The extracted percussive sounds and drum track serves as a basis for search by rhythmic similarity in the context of the European project Cuidado.

Pulse-dependent analyses of percussive music

We report on a method of automatic extraction of a metrical attribute from percussive music audio signals: the smallest rhythmic pulse, called the ìtickî. The relevance of use of this feature in the framework of subsequent analyses is discussed and evaluated.Learning Bayesian Belief Networks (BBN) from corpora and incorporating the extracted inferring knowledge with a Support Vector Machines (SVM) classifier has been applied to character segmentation for unconstrained handwritten text. By taking advantage of the plethora in unlabeled data found in image databases in addition to some available labeled examples, we overcome the expensive task of annotating the whole set of training data and the performance of the character segmentation learner is increased. Apart from this approach, which has not previously used for this task, we have experimented with two well-known machine learning methods (Learning Vector Quantization and a simplified version of the Transformation-Based Learning theory). We argue that a classifier generated from BBN and SVM is well suited for learning to identify the correct segment boundaries. Empirical results will support this claim. Performance has been methodically evaluated using both English and Modem Greek corpora in order to determine the unbiased behaviour of the trained models. Limited training data are proved to endow with satisfactory results. We have been able to achieve precision exceeding 86%.

MTG-DB: a repository for music audio processing

Content-based audio processing researchers need audio and its related metadata to develop and test algorithms. We present a common repository of audio, metadata, ontologies and algorithms. We detail the hardware implementation, in the form of massive storage and computation cluster, the software and databases design and the ontology management of the current system. The repository, as far as copyright licenses allow, is open to researchers outside the music technology group to test and evaluate their algorithms.

A beat induction method for musical audio signals

We present a method for segmenting musical audio signals with respect to a particular metrical level: the beat. No assumption has been made regarding sound sources. We situate this proposal with respect to recent models. The model proposed seeks recurrences in values of audio signal low-level features. These features are computed at the scale of the smallest metrical level: the tick —or tatum. Focusing on energy features in frequency subbands gave better results than on the whole frequency range. An attractive aspect of the method is that it permits to evaluate the relevance of any low-level feature as a cue for beat induction.