Frank Seifert

Semantic relationship and identification of music

Automatic search engines for musical documents mostly provide a vast number of relevant and irrelevant hits including multiple appearances of the same documents. Generally, underlying information-processing methods that are based on metadata or simple content-based procedures cause this situation. Therefore we propose a model for the recognition of music that enables identification and comparison of musical documents regardless their actual instantiation. Inherent component of this model is the capability of navigation based on semantic relationship between musical pieces. In order to realize this functionality we introduce a concept that tries to derive deductive music recognition from generic music production as an equivalent counterpart. Hence, some deductive processes of human music perception are simulated. The resulting structures enclose musical meaning and therefore can be used for the estimation of identity and relationship between musical documents. As a byproduct of this functionality, plagiarism and copyright infringements could be detected.

Semantic music recognition - audio identification beyond fingerprinting

Besides audio fingerprinting techniques there are no essential procedures for a content-based identification of music audio available. But even these techniques rely heavily on statistical information of audio and do not consider any semantics of music. Furthermore, they require each piece of music to be pre-recorded and thus pre-processed for a successful identification. We try to apply the leadsheet-model - a generic model for processing tonal music - on content-based audio identification and show how it can be altered to handle audio. As a result we are capable of identifying music with extremely varying spectra based on only one given template.

Melodic Segmentation on Different Musical Genres

To create a sufficient repository of test data for our model-based implementations of music information retrieval functions working on symbolic documents, we have used three different approaches to melodic segmentation on monophonic pieces with a broad range of genres from baroque/classical music to pop/rock. All three methods are driven by musical knowledge (in contrast to methods such as n-gram segmentation). Two of the algorithms we have applied are taken from former work of other researchers, the third algorithm has been developed in our department and will be introduced briefly. Our repository of test documents (midi format) consisted of 52 files making it more representative (2.5 to 5 times the number of documents)than those that have been referenced in previous publications. This paper describes our experiences with the applied algorithms, the results that have been achieved, and the conclusions we have been able to draw for improving music segmentation methods.

Pattern-based Classification of Rhythms

We present a pattern-based approach for the rhythm classification task that combines an Auto-correlation function (ACF) and Discrete Fourier transform (DFT). Rhythm hypotheses are first extracted from symbolic input data, e.g. MIDI data, by the ACF. These hypotheses are validated by the use of DFT to remove duplicates before the classification process. The classification of rhythms is performed using ACF in combination with rhythm patterns contained in a knowledge base. In this, the input data is classified against basic rhythmic patterns in the knowledge base. The evaluation of this method is performed using pre-labelled input data. We show that a knowledge-based approach is reasonable to address the problem of rhythm classification for symbolic data.

Generic modeling of music for computational education and recognition of song variations

For purposes of identification and education todays music modeling is rather limited. Systems that can identify music in a generic way are either very restricted or do not exist at all. Therefore, it is not possible to associate the countless potential occurrences of a certain song with at least one generic description. Analogously, there is no computational approach to evaluate the performance of a music student who has been asked to play a certain song and to accompany it in a certain style. We introduce a generic model for music representation. To test our model, we have developed a prototype that correlates symbolic music performances to the proposed representation.

Generic Music Identification by Hierarchic Modeling of Human Perception

To date, there are no systems that can identify music, neither audio nor midi, in a generic way. That is, for the countless possible occurrences of a certain song, it should be possible to associate them with only one generic description of that song. In this paper we describe a conceptual design for generic music identification. Therefore, we propose a hierarchy of elementary music descriptions and introduce the concept of generic perceptual musical patterns. To prove the designed model, it is implemented for symbolic music files.