Howling Corrupted Music Speech (HCMS)

The database contains in total 28 music files and 30 speech files. The original speech files are taken from 8 original signals files including recorded male and female speech in four languages (Chinese, English, Dutch, and Russian) from an audiobook database [1]. while the original music files is taken from 7 pieces spanning various genres (e.g., jazz, opera) from 2 different music databases [2, 3]. Each file is a 20 s excerpt where the howling is simulated to start between the 8th and 9th second. This is simulated by feeding the music or speech source signal to a closed-loop system with a varying broadband gain. A total of 8 acoustic impulse responses (AIRs) were used for the simulations, hence covering a wide range of howling frequencies. The final dataset1 contains fewer excerpts than the number of excerpts originally generated, as it went through a pruning step conducted by three experts, aiming to eliminate unsuitable examples (exhibiting howling at multiple howling frequencies or no howling at all). For more details see Ch. 7 of" Acoustic Event Detection: Feature, Evaluation and Dataset design", PhD thesis Mina Mounir.

[1] Kearns, J.: LibriVox: Free Public Domain Audiobooks. Emerald Group Publishing
Limited (2014)
[2] Emiya, V.: MAPS Database: A piano database for multipitch estimation and automatic transcription of music. http://www.tsi.telecom-paristech.fr/aao/en/2
010/07/08/maps-database-a-piano-database-for-multipitch-estimation-and-aut
omatic-transcription-of-music/ (2008)
[3] Defferrard, M., Benzi, K., Vandergheynst, P., Bresson, X.: FMA: A DATASET
FOR MUSIC ANALYSIS. Proc. 18th Int. Symp. on Music Information Retrieval
(ISMIR ’17), 8 (2017)