Systèmes de Recherche de Documents Musicaux par Chantonnement

Abstract : With the explosion of the digital data available (via Internet particularly), a growing attention has been paid to the means of access to documents. Document indexing, traditionally based on text, provides imprecise descriptions, especially for musical contents.

This thesis focuses on the retrieval of musical documents by humming. We propose a system that performs a search for music initiated from a short excerpt, hummed by the user. This work includes two preliminary studies that forefill some missing elements of a still young research domain. We consider the accuracy of hummed melodies observing 500 queries. We also reveal some aspects of melodic similarity by carrying out subjective tests.

Thanks to these studies, we propose a both original and efficient retrieval system. Refusing a tempered description of the query (i.e. using a quantization of notes into semi-tones), our system retrieves more than 90% of the expected documents from a mean length query (13 notes). The database is made of 20,000 MIDI files (40 millions indexed notes). The search time is acceptable for common use, as the system takes only a few seconds to return the result (i.e. the list of documents that are the most similar to the query).

This thesis also contributes to the quality evaluation process of such systems. We propose a model of the imprecision of hummed melodies. This allows artificial queries to be constructed and substitued to real queries for system testing. This makes the test process easier while preserving a realistic stimulation of tested systems.
Complete list of metadatas

Cited literature [59 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-00001593
Contributor : Matthieu Carré <>
Submitted on : Friday, August 30, 2002 - 12:40:39 PM
Last modification on : Wednesday, February 20, 2019 - 2:41:49 PM
Long-term archiving on : Friday, April 2, 2010 - 6:24:13 PM

Identifiers

  • HAL Id : tel-00001593, version 1

Citation

Matthieu Carré. Systèmes de Recherche de Documents Musicaux par Chantonnement. Interface homme-machine [cs.HC]. Ecole nationale supérieure des telecommunications - ENST, 2002. Français. ⟨tel-00001593⟩

Share

Metrics

Record views

296

Files downloads

533