Integration of neural networks and probabilistic spatial models for acoustic blind source separation / von M. Sc. Lukas Drude ; Erster Gutachter: Prof. Dr.-Ing. Reinhold Häb-Umbach, Zweiter Gutachter: Prof. Dr.-Ing. Timo Gerkmann. Paderborn, 2020
Inhalt
- Abstract
- Acknowledgments
- Introduction
- Prerequisites
- Notation
- Signal model
- Overview table of variable names
- Random variables
- Latent variable models and the expectation maximization algorithm
- Blind source separation principles
- Principles of single-channel approaches
- Principles of multi-channel approaches
- Probabilistic spatial mixture models
- Frequency permutation problem
- Initialization
- Influence of the mixture weight
- Complex Watson mixture model
- Complex Bingham mixture model
- Full-Bayesian complex Watson mixture model
- Time-variant complex Gaussian mixture model
- Complex angular central Gaussian mixture model
- Guided source separation
- Spatial features for neural networks
- Principles of source extraction
- Integration of neural networks and probabilistic graphical models
- Existing integration approaches
- Cascade approach: Integration by initialization
- Tight integration of spatial and spectral features
- Unsupervised training using multi-channel features
- Evaluation
- Performance metrics
- Database design
- Acoustic model training
- Deep-learning methods
- Deep clustering
- Deep attractor network
- Permutation invariant training
- Comparison with reference publications on WSJ0-2mix
- Probabilistic spatial mixture models
- Source extraction
- Integration of neural networks and probabilistic graphical models
- Weak integration: A cascade approach
- Strong integration
- Comparison of integration models with single-/ multi-channel encoder
- Unsupervised training of deep clustering
- Overview of all methods on WSJ-BSS
- Overview of all methods on WSJ-MC
- Reproducibility and statistical significance
- Conclusion
- Appendix
- Properties of the complex Bingham distribution
- Non-negativity of the Kullback-Leibler divergence
- Mixture weights without Lagrange's method
- Remarks on complex derivatives
- GEV/MaxSNR beamformer
- MVDR beamformer
- Permutation formalism
- Comparison of WSJ-BSS and SMS-WSJ
- More detailed evaluation results
- Glossary
- List of peer-reviewed publications with own contributions (OC)
- Bibliography