Sung Note Segmentation for a Query-by-Humming System
Pradeep Kumar, M. V. Joshi, S. Dutta-Roy, Preeti Rao · 2007
Abstract. Retrieval performance in query-by-humming (QBH) systems depends crucially on the accurate note segmentation and labeling of user queries. To facilitate note segmentation, querying is often restricted to the easily detected syllable “ta”, which is not necessarily the syllable most preferred by users. In this work, new acoustic features based on the signal energy distribution as obtained from the singing perception and production points of view are investigated. Performance evaluations on a manually labeled database of syllabic humming show that a specific mid-band energy combined with a biphasic detection function achieves high correct detection and low false alarm rates on the sonorant consonant syllables /da/, /la / and /na/. The resulting onset detector is incorporated in the signal-processing front-end of an available QBH system (hitherto constrained to ta-syllable queries only). QBH retrieval performance results are reported on a large dataset of user queries.