Abschlussarbeit – Verfahren zur Erkennung akustischer Ereignisse

This challenge awaits you :

With new developments in the field of AI and machine learning, high-performance methods are increasingly being used for the automatic processing, analysis, and generation of audio data. Like this, Machines can extract information from audio signals, like a human who continuously perceives information in their acoustic environment. Among the various applications, unwanted noises in vehicles can be identified, glass breakage events in urban environments can be detected, or animal sounds can be classified.

For reliable detection of acoustic events using supervised learning methods, a significant effort is still required for the collection, analysis, and labeling of training data. Generally, the shorter and rarer an acoustic event is, the fewer recordings are available in which the event is not only present but also correctly labeled. Solutions to this challenge can be achieved by better utilizing the exponentially growing amount of available data from all sources or by reducing the need for large (initial) training data sets with a machine learning method optimized for this purpose. From the following topics in this context can be chosen :

Semi-Supervised Learning :

It can be assumed that for particularly difficult-to-label acoustic events, a larger amount of data exists that contains the sound but lacks a label or has an incorrect label. Semi-supervised learning methods can generate reliable models for detection and classification from a limited amount of labeled data and a large amount of unlabeled data.

Incremental Learning :

With incremental learning, a system can continuously learn and adapt to new acoustic events without the need for a complete retraining. The amount of initial training data can also be kept feasible. The increments with additional training data can be obtained through continuous inference of the models on a large, unlabeled data set, allowing the system to acquire new training data itself.

Few-Shot Learning :

The use of few-shot learning techniques enables the training of models that can deliver precise results with only a few training examples. In this thesis, a method should be developed to detect and classify vehicle-related acoustic events using few-shot learning with a minimal data set.

This thesis offers the opportunity to work at the intersection of machine learning and acoustic signal processing. The goal is to explore one of the mentioned methods for acoustic event detection and enable the use of larger training data sets without additional labeling effort or achieve increased accuracy while using a small available (training) data set.

Your Tasks :

You conduct literature research and familiarize yourself with the topic

You develop and implement methods for detecting acoustic events or anomalies

You analyze and optimize existing models for detecting and classifying acoustic events

You conduct experiments and tests to validate the developed models

You document and present your work results

Necessary Skills :

• Ongoing studies in computer science, mechatronics, electrical engineering, or a comparable technical or scientific course of study
• Programming experience in at least one high-level programming language
• Excellent English skills, both written and spoken
• Good German skills (Level B2 or higher)

Desired Skills :

Remuneration is based on our collective wage and salary agreement. The current monthly salary for this position is EUR 979.00.

You won’t just be working anywhere as a student at IAV. You’ll be right in the middle of it all. Real projects. Exciting future tasks. Completely integrated and side-by-side with IAV experts. Lots of responsibility and at the same time lots of freedom, combining university and work. The result is the best prospects for your professional development. And attractive compensation in accordance with our company wage agreements.

Diversity and equal opportunity are important to us. What matters to us is the individual, with his or her character and strengths.