Deep Learning for Music Unmixing

Problem Statement

Funny slide

Classical Approaches: Unsupervised

• NMF
• KAM
• short intro into source-filter models, NMF, repet, kam...
• limitations

Paradigna shift

• when?
• why? data, separation is only well defined for certain kind of data

Supervised Source Separation

Supervised Source Separation

• Classification vs Regression Binary Masking, Deep Clustering

• Discriminate vs Generative

• Single I/O vs Multiple I/O Siamese networks, Deep-U net

• Time Series vs. Frequency

• its not clear if its a discriminate or generative problem

• permutation problem

Baseline System (Popular Decisions)

• BiLSTM - Sequence to Sequence
• PreProcessing
• Structure
• Training
• Testing
• Explain UHL2 (BiLSTM) == state of the art
• deep clustering

SiSEC2018 Demotime

How hip is Source separation

• End-to-End Networks
• Fully Convolutional Networks
• Batch Normalization
• GANs
• Capsule Networks
• Attention Networks
• Reinforcement Learning

Research Directions

• Unsupervised Learning to solve the permutation problem
• Learning with limited amount of data
• Improving Representations vs End-to-End (WaveNet)
• datasets accessible for non-audio experts, require significant domain knowledge
• Convergence of Signal Processing and DL
• Model long term dependency