Using FFT Magnitudes and ELM for Multi-pitch Estimation

摘要：

This work proposes and investigates a system for multi-pitch estimation of musical signal using extreme learning machine (ELM) with features generated from FFT magnitudes. We reduce a numbers of features using frequency space scaling and compare evaluation results between some variations of this method. The ELM in our system performs in multioutput regression mode to map between input features and outputs of instrument notes. Multi-pitch estimation can be seen as multi-label classification. However, selecting class based on the highest output is impossible since multiple outputs can be active simultaneously. We therefore rely on threshold function to reconstruct the ELM outputs instead. From the evaluations, our system gives the best results when using 24-tet frequency space scaling with nonoverlap selective window as a dimensional reduction method. This is also better than using PCA even the sizes of the final reduced features from both methods are the same.

关键词： Multi-pitch estimation ELM multi-label classification

作者: Pat Taweewat

作者单位: School of Electrical and Information Engineering The University of Sydney NSW, Australia

会议类型: 国际会议

会议名称: 2010 International Conference on Measurement and Control Engineering(2010年IEEE测量与控制工程国际会议 ICMCE2010)

会议地点: 成都

会议语种:英文

页码: 402-406

在线出版日期: 2010-11-16（万方平台首次上网日期，不代表论文的发表时间）