Die Veranstaltung soll ein breites und hinreichendes tiefes theoretisches Fundament legen, um in Wissenschaft und Praxis mit Cochlea Implantaten (CIs) und CI Trägern arbeiten zu können.
Understanding the signal processing principles applied to
hearing devices (hearing aids and cochlear implants)
Contents:
- Amplification and compression
- Speech enhancement and noise reduction
- Signal processing in cochlear implants
- Computational auditory scene analysis
- Automatic classification of the acoustic environment
- Acoustic feedback management
The field of Machine Learning develops and provides methods for the analysis of data and signals. Typical application domains are computer hearing, computer vision, general pattern recognition and large-scale data analysis (recently often termed "Big Data"). Furthermore, Machine Learning methods serve as models for information processing and learning in humans and animals, and are often considered as part of artificial intelligence approaches.
This course gives an introduction to unsupervised learning methods, i.e., methods that extract knowledge from data without the requirement of explicit knowledge about individual data points. We will introduce a common probabilistic framework for learning and a methodology to derive learning algorithms for different types of tasks. Examples that are derived are algorithms for clustering, classification, component extraction, feature learning, blind source separation and dimensionality reduction. Relations to neural network models and learning in biological systems will be discussed were appropriate.
The course requires some programming skills, preferably in Matlab or Python. Further requirements are typical mathematical / analytical skills that are taught as part of Bachelor degrees in Physics, Mathematics, Statistics, Computer and Engineering Sciences. Course assignments will include analytical tasks and programming task which can be worked out in small groups.
The presented approach to unsupervised learning relies on Bayes' theorem and is therefore sometimes referred to as a Bayesian approach. It has many interesting relations to physics (e.g., statistical physics), statistics and mathematics (analysis, probability theory, stochastic) but the course's content will be developed independently of detailed prior knowledge in these fields.
Weblink: www.uni-oldenburg.de/ml
The students will gain in-depth knowledge on the subjects’ speech and audio processing. The practical part of the course mediates insight about important properties of the methods treated in a self-study approach, while the application and transfer of theoretical concepts to practical applications is gained by implementing algorithms on a computer.
content:
After reviewing the basic principles of speech processing and statistical signal processing (adaptive filtering, estimation theory), this course covers techniques and underlying algorithms that are essential in many modern-day speech communication and audio processing systems (e.g. mobile phones, hearing aids, headphones): acoustic echo and feedback cancellation, noise reduction, dereverberation, microphone and loudspeaker array processing, active noise control. During the exercises a typical hands-free speech communication or audio processing system is implemented (in Matlab).
The students will learn about recent developments and state-of-the-art approaches in Machine Learning, and their applications to different data domains. By presenting scientific studies in the context of currently used models and their applications, they will learn to understand and communicate recent scientific results. The presentations will use computers and projectors. Programming examples and animations will be used to support the interactive component of the presentations. In scientific discussions of the presented and related work, the students will obtain knowledge about current limitations of Machine Learning approaches both on the theoretical side and on the side of their technical and practical realizations. Presentations of interdisciplinary research will enable the students to carry over their Machine Learning knowledge to address questions in other scientific domains.
Contents:
In this seminar recent developments of models and algorithms in Machine Learning will be studied. Advances of established modelling approaches and new approaches will be presented and discussed along with the applications of different current algorithms to application domains including: auditory and visual signal enhancements, source separation, auditory and visual object learning and recognition, auditory scene analysis and inpainting. Furthermore, Machine Learning approaches as models for neural data processing will be discussed and related to current questions in Computational Neuroscience.
Exercises
Hinweise zum Modul
Prerequisites
Bachelor in Hörtechnik und Audiologie oder entsprechend
Further information
Module Capacity: unrestricted
Method of assesment
M
Learning outcomes/competencies
Vermittlung der theoretischen Grundlagen und praktischen Anwendungen moderner Sprachtechnologie. Vermittlung moderner Signalverarbeitungsalgorithmen für digitale Hörgeräte, Cochlear Implantate, Sprachkommunikations- und Audiosysteme. Vermittlung der Grundlagen der Informationsverarbeitung und Informationstheorie, und praktischer Methoden der statistischen Signalverarbeitung, Signalkompression und Nachrichtenübertragung. Messungen akustischer Ereignisse sowie Messungen zur Identifizierung akustischer Systeme. Nach Abschluss des Moduls beherrschen Studierende (a) moderne Signal- und Informations-verarbeitungsmethoden und können (b) die gelernten Methoden zur Analyse schwingungsphysikalischer Systeme und zur Erklärung der Funktionsweise und Analyse signalverarbeitender Systeme einsetzen.