Machine Learning for Speech and Audio Processing

Lecturer: Prof. Dr.-Ing. Peter Jax

Contact: Lars Thieling, Maximilian Kentgens

Type: Master lecture

Credits: 4

Lecture in RWTHonline
Exercise in RWTHonline
Learning room RWTHmoodle
(Registration via RWTHonline)

Course language: English

Material:
Lecture slides and Exercise problems will be published in RWTHmoodle.

Dates

Lecture:

from Friday, April 16, 2021
08:30 - 10:00
online

Exercise:

from Friday, April 16, 2021
10:15 - 11:00
online

Exam

Thursday, August 5, 2021
17:00-18:30
TEMP 1 / TEMP 2

Exam duration: 90 minutes

Remarks: The exam is in written form. The date corresponds also to the proof of achievements (Leistungsnachweise) in written form.

Resources: You are allowed to bring one hand-written DIN A4 formula sheet (front and back). Any other written material (e.g., lecture notes, exercise notes) is not allowed. A non-programmable calculator is allowed.

Please note: Please bring along your student ID (BlueCard)!

The lecture "Machine Learning for Speech and Audio Processing (MLSAP)" addresses especially students of the Master's program "Electrical Engineering, Information Technology and Computer Engineering". The formal connection to the module catalogs can be found at RWTHonline.

Content

In this one term lecture the fundamental methods of machine learning with applications to problems in speech and audio signal processing are presented:

  • Fundamentals of Classification and Estimation
    • Basic Problems of Classification
    • Feature Extraction Techniques
    • Basic Classification Schemes
  • Probabilistic Models
    • Stochastic Processes and Models
    • Gaussian Mixture Models (GMMs)
    • Hidden Markov Models (HMMs)
    • Training Methods
    • Bayesian Probability Theory: Classification and Estimation
    • Particle Filter
  • Non-Negative Matrix Factorization (NMF)
    • Dictionary-based concept
  • Neural Network and Deep Learning
    • Feed-Forward Neural Networks
    • Fundamental Applications
    • Learning Strategies: Supervised vs Unsupervised vs Reinforcement Learning
    • Training of Synaptic Weights: Backpropagation and Stochastic Gradient Descent
    • Behavior of Learning and the “Magic” of Setting Hyper‐Parameters
    • Generative Networks as a Complement to Directed Graphs
    • From „Shallow“ to „Deep“: Trade Comprehensibility for Performance
    • Specific Network Architectures
    • Applications in Signal Processsing
    • Interpretations and Realizations

Exercises are offered to gain a deeper understanding on the basis of practical examples.

Evaluation

The results of the evaluation are summarized below.

Summer term 2020

Participants of the evaluation (lecture/exercise): 16/10

Lecture:
Global grade: 1,4
Concept of the lecture: 1,4
Instruction and behaviour: 1,4

Exercise:
Global grade: 1,4
Concept of the exercise: 1,5
Instruction and behaviour: 1,4

MLSAP_Vorlesung_SS20.pdf
MLAP_Uebung_SS20.pdf