Single-Channel Source Separation Tutorial Mini-Series

by Nicholas Bryan, Dennis Sun, and Eunjoon Cho


Lecture 1: Classical Speech Denoising and Enhancement

Abstract: To start off a series of three tutorial-style dsp seminars on current single-channel source separation methods, the first talk will introduce the topic of classical methods used for speech enhancement. We will cover denoising methods such as spectral subtraction, wiener filters, and probabilistic estimators (if time permits). This will give an overview of the large range of literature on reducing background noise from speech.

Given the nature of the application, most of these methods have to work in real-time settings and have to be robust to non-stationary noisy environments. We will discuss some of the advantages and shortcomings of these methods in real-world scenarios when it comes to enhancing noisy speech. Following the lecture, we will also get a chance to program some of these speech enhancement algorithms and see how they perform under various settings. Please bring your laptops with Matlab and/or Octave installed and be ready to code!

Speaker Bio: Eunjoon Cho is a PhD candidate in the Department of Electrical Engineering at Stanford. His research focus is on using the underlying structure of speech to estimate background noise in non-stationary environments.

Materials: Slides I Code I

Lecture 2: Introduction to Non-Negative Matrix Factorization

Abstract: For the second talk in this series, we will introduce the topic of non-negative matrix factorization for the purpose of single-channel source separation. NMF is one of the current most promising and effective class of approaches found for source separation and is a popular topic in several signal processing conferences and journals. Following the lecture, we will get a chance to program a basic source separator. Please bring your laptops with Matlab and/or Octave installed and be ready to code!

Speaker Bios: Nicholas J. Bryan is a PhD candidate at the Center for Computer Research in Music and Acoustics (CCRMA) working with Prof. Ge Wang. His research interests are at the intersection between signal processing, machine learning, and human-computer interaction.

Dennis Sun is a PhD candidate in the Department of Statistics and working with Prof. Jonathan Taylor. His research interests are at the interface of statistics, signal processing, machine learning, and musicology.

Materials: Slides II Code II

Lecture 3: Extensions and Interpretations to Non-Negative Matrix Factorization

Abstract: Building off the last two lectures in the series, we will continue our discussion on non-negative matrix factorization techniques for source separation. We will talk about common extensions, additional interpretations, methods of evaluation, and if time permits, future directions of research. Following the lecture, we will get a chance to program and improve our basic separator from the second lecture. Please bring your laptops with Matlab and/or Octave installed and be ready to code!

Speaker Bios: Dennis Sun is a PhD candidate in the Department of Statistics and working with Prof. Jonathan Taylor. His research interests are at the interface of statistics, signal processing, machine learning, and musicology.

Nicholas J. Bryan is a PhD candidate at the Center for Computer Research in Music and Acoustics (CCRMA) working with Prof. Ge Wang. His research interests are at the intersection between signal processing, machine learning, and human-computer interaction.

Materials: Slides III Code III