We offer student research projects, and bachelor's and master's theses in the following research areas. Please contact us!
Contact persons for these and other topics are Prof. Dr.-Ing. Rainer Martin and the scientists named on the respective topics.
Student theses can be written in German or English. Teamwork is possible for practical projects (Bachelor's) and Master's projects.
What can you learn?
In addition to an insight into current issues and results from research, student theses offer the opportunity to apply the knowledge acquired during your studies. Research, analysis, and dealing with scientific literature are trained as well as structured problem-solving, programming, and the use of relevant software tools and/or hardware, e.g. for acoustic measurements or the training of neural networks.
What should you bring with you?
In addition to an interest in issues relating to speech and audio signal processing, you should above all have a solid basic knowledge of systems theory and signal processing, as taught for example in the courses Systems Theory 1 to 3 and Speech and Audio Communication. Further work can also build on the content of Master's courses (communication acoustics, digital signal processing). You should have a basic knowledge of English that allows you to study scientific literature.
What is it about?
Whether you're on the phone on the train, listening to music in the canteen, or using a hearing aid: In all of these situations, ambient noise impairs the perception of the acoustic signal. This interference can be alleviated using noise reduction and speech signal enhancement methods. While noise reduction aims to improve the perceptibility of the desired signal, such as music, by reducing background noise and distortion, speech signal enhancement also aims to improve speech intelligibility.
Solution approaches
Speech enhancement methods often use statistical estimation techniques, more recently often in combination with machine learning, to identify speech and interference in the captured signal. The aim is to extract and reproduce the speech signal with minimal distortion. Active noise control (ANC), also known as active noise cancellation, considers possible interference from noise during playback. By precisely modeling the transmission paths and estimating the interference signals, counter-signals can be introduced to cancel out the noise to a large extent by destructively interfering with the acoustic signals. If several microphones are available on the device used, so-called beamforming algorithms are a means to utilize directional information in the signals, which benefits the functionality of both methods.
Both approaches use analytical or numerical optimization methods that minimize a cost function based on the problem. Nowadays, deep neural networks (DNNs) are also frequently used. For use in real-time systems, minimizing the latency between the input of the disturbed signal and the output of the processed improved signal is essential.
Specifically, this requires that filter structures should not be too complex, and algorithms should deliver good results even with very short signal segment lengths (e.g., 20 ms).
Tools and methods
The Bachelor's and Master's theses that we offer focus on improving existing algorithms and implementing and evaluating new approaches. The problems are solved using audio signal processing algorithms, statistical signal processing, and machine learning. The corresponding implementations are typically realized in MATLAB or Python. Depending on the task, simulated results can be complemented by real-time measurements in our laboratories.
What will you learn?
As part of a Bachelor's/Master's thesis, you will acquire basic knowledge of audio signal processing, acoustics, and statistical signal processing. In addition, programming skills are advanced through the efficient implementation of algorithms.
Contact:
Willem Alexander Klatt, M. Sc.
ID 2/319
willem.klatt@rub.de
Benjamin Lentz, M. Sc.
ID 2/319
benjamin.lentz@rub.de
Maurice Oberhag, M. Sc.
ID 2/319
maurice.oberhag@rub.de
What is it about?
In many voice communication applications, several voice signals are recorded simultaneously, of which only one is usually desired, while the others are experienced as disturbing and should therefore be suppressed. For example, a hearing aid user wants to listen to the person they are talking to at a party, even though the microphones of their hearing aids also pick up the speech of the other guests; a speech assistance system should understand commands and requests as accurately as possible, even if children are playing or television is playing in the background; and when communicating hands-free with a cell phone, the person I am talking to should primarily hear my voice and not their own echo, which is fed back into the microphone by the loudspeaker.
In all these cases, signal processing can be used in conjunction with machine learning methods to separate the desired speech signal from the other speech signals. In student projects, students develop and implement appropriate methods from the scientific literature or from current research at the chair, and then evaluate and further develop them.
Tools and methods
The tasks are mainly worked on in Matlab / Simulink, although development in e.g. Python can also be useful in some work. Methodologically, digital and statistical signal processing methods (e.g. fast convolution, estimation methods) form the basis. In addition, methods from the field of machine learning (e.g. neural networks, deep learning) or adaptive filters can be used.
Examples of completed work on speaker separation:
Contact:
Prof. Dr.-Ing. Rainer Martin
ID 2/233
0234 32-22495
rainer.martin@rub.de
What is it about?
Localizing, tracking, and recognizing sources in complex acoustic scenes is one of the most challenging tasks in acoustics. A key objective is to determine the position of a source or several sources relative to the recording microphones. Applications can be found, for example, in robotics or in hearing aids. Microphone groups (“arrays”) are generally used to solve the task. The information obtained can then be used for beamforming, for example, which focuses on the source signals and suppresses sources of interference.
An acoustic scene classifier assigns the signals recorded by the microphones to a signal class or a predefined acoustic scene. Examples of acoustic signal classes include noise, music, or speech. This information can be used, for example, to control the audio signal processing in hearing aids according to the recognized acoustic signal class. If, for example, the noise class is detected, the noise reduction algorithm in the hearing aid is automatically activated.
An acoustic scene is the time-varying interplay of spatially distributed sources found in certain environments. For example, the scene “office environment” is characterized by voices, telephone ringing, keyboard noises, and printer noises. In the analysis, the scene is then determined by information about the temporal and spatial characteristics of the recorded microphone signals.
The Bachelor's/Master's thesis focuses primarily on the improvement of existing algorithms and the development and implementation of new algorithms on various platforms (PC, Android, real-time systems).
Tools and methods
The problems are solved with algorithms for audio signal processing, statistical signal processing, and machine learning methods. The implementation is carried out using Matlab, Simulink, or Python.
What will you learn?
As part of a Bachelor's/Master's thesis, you will learn the basics of pattern recognition, audio signal processing, and statistical signal processing. In addition, programming skills and the efficient implementation of algorithms are deepened.
What criteria do you need to meet?
Willingness to acquire basic knowledge of acoustics, audio signal processing, and machine learning. Programming skills in Matlab and Simulink, knowledge of C++ and Python are an advantage, as well as an interest in recording technology and electronics.
Examples of completed work:
Contact:
Prof. Dr.-Ing. Rainer Martin
ID 2/233
0234 32-22495
rainer.martin@rub.de
What is it about?
Acoustic Sensor Networks (ASNs) and in particular Wireless ASNs (WASNs) are considered the next-generation technology for audio signal recording and processing. Compared to traditional microphone arrays, ASNs offer both more flexibility, as they can be integrated into a wide variety of embedded systems, and better scalability, as they can cover small and large acoustic scenarios alike.
Some examples are assisted living and smart homes, where sensors are distributed in a room or throughout the house, to entire smart cities or the detection and monitoring of the natural habitat of animals, with many sensors distributed over a wide area.
The technical tasks and challenges that these networks address vary from speaker and speech recognition to the detection of acoustic events and the grouping and classification of sound sources. There are many examples of different so-called classification labels: ‘car horn’, ‘siren’, ‘telephone ringing’, ‘washing machine running’, ‘chainsaw cutting’, ‘engine noise’, and many more.
What you can expect
The main environments used are Matlab and/or Python, additionally any experience with Tensorflow is welcome. Some projects have also been developed with and on Android devices in Java.
You will have the opportunity to use the most important signal processing tools (FFT, DCT, MFCC, etc.) and combine them with machine learning approaches (LDA, PCA, clustering, entropy, mutual information, etc.). During some projects you will also have the opportunity to learn more about and use state-of-the-art deep neural network technologies (CNN, MLP, GRU, etc.).
The projects can be worked on individually or in a team. In addition, you will learn helpful basic techniques in the area of agile project management, such as sprints and scrums.
Examples of work already completed:
Bachelor practice projects:
Bachelor's theses:
Contact:
Luca Becker, M.Sc.
ID 2/255
0234 / 32-27543
luca.becker@rub.de
What is it about?
While signal processing for hearing aids and cochlear implants (CI) has long been focused exclusively on improving the quality of speech signals, other types of signals, such as music, are becoming increasingly important to developers and users. Music signals differ from speech signals in several ways: their composition is much more varied and complex, and they cover a much wider frequency and dynamic range.
Users of hearing aids and CIs, on the other hand, often suffer from effects such as a severe reduction in their spectral resolution, spectral smearing, and a reduction in the volume range perceived as pleasant due to their hearing impairment. They have difficulty hearing individual voices or instruments in pieces of music. These impairments often make listening to music more strenuous for them than for people with normal hearing.
The goal of our work is to make music signals more accessible for hearing aid and CI users, e.g. by amplifying or attenuating individual voices, instruments, or rhythmic elements, or by simplifying or otherwise processing the spectrum of the signals. This is done using techniques from the fields of source separation, music information retrieval, digital effects, statistical signal processing, machine learning, and neural networks. As a rule, it is necessary to test the newly developed algorithms for their effect and applicability in hearing tests with normal hearing, hearing aid, and CI listeners.
Previous work:
Contact:
Dr.-Ing. Anil Nagathil
ID 2/223
0234 32-29289
anil.nagathil@rub.de
What is it about?
"Virtual reality" can be described as a plausible visual and auditory representation of a virtual environment. Visual stimuli can be displayed using VR glasses, while acoustic stimuli are reproduced using binaural (two-ear) headphones and loudspeaker systems. The quality of a virtual environment is characterized by how immersive it is, i.e. how well you can immerse yourself in it and perceive it as realistic. The better this is achieved, the better the integration of reality and virtual reality. This requires that both the visual and acoustic environments are coherent and in harmony with each other.
Tasks
In the student projects, sound transmission from the source to the binaural receiver is modelled using the so-called Head-Related Transfer Function (HRTF). This can be measured individually for each person in the HRTF laboratory at the Institute of Communication Acoustics. With comprehensive knowledge of the HRTF, binaural signals from all conceivable spatial directions can then be synthesised and examined for their quality. This can be done for the simplest possible (anechoic) scenarios, such as in an open field, as well as for more complex scenarios (rooms; multiple sources). The generated virtual acoustic environment is then combined with a virtual visual environment (e.g. in Unity). This created virtual reality can then be evaluated in various applications, e.g. gaming, by checking whether objects can be easily localised within this virtual reality.
Tools and methods
The metrological side of the task includes the operation of equipment and audio recording and calculation software (MATLAB) in the HRTF laboratory. Binaural signals are synthesised in MATLAB or in the application-oriented Unity 3D environment with C++/C#. The latter is particularly suitable for adding a visual modality to the acoustic modality.
What will you learn?
Depending on the focus of the respective Bachelor's/Master's thesis, you will learn the basics of signal acquisition or signal synthesis. Algorithmically, you will deepen your knowledge of the methods of adaptive signal processing and machine learning. Programming skills for the efficient implementation of algorithms and applications are trained.
What criteria do you need to meet?
Basic knowledge of audio signal processing, such as acquired in the Speech and Audio Communication course. In-depth knowledge of digital and adaptive signal processing, such as from the Master's degree course in Communication Technology. Programming skills in Matlab, C or Python.
Contact:
Daniel Neudek, M. Sc.
ID 2 / 221
0234 / 32 25388
daniel.neudek@rub.de
Copyright © IKA 2024
Last update: Nov 20, 2024