Abstract

A music retrieval system that matches a short length music query with its variations in a database is proposed. In order to avoid the negative effects of different orchestration and performance style and tempo on transcription and matching, a mid-level representation schema and a tonal modeling approach is used. The mid-level representation approach transcribes the music pieces into a sequence of music tags corresponding to major and minor triad labels. From the transcribed sequence, n-gram models are built to statistically represent the harmonic progression. For retrieval, a perplexity based similarity score is calculated between each n-gram in the database and that for the query. The retrieval performance of the system is presented for a dataset of 2000 classical music pieces modeled using n-grams of sizes 2 through 6. We observe improvements in retrieval performance with increasing query length and n-gram order. The improvement converges to a little over one for all query lengths tested when n reaches 6.

Date

2011

Authors

Erdem Unal, Elaine Chew, Panayiotis G Georgiou, Shrikanth S Narayanan

Conference

ISMIR

Pages

43-48