Directly to the content Skip to maincontent Directly to the footer Skip to footercontent

Master's thesis: "Development of a machine learning approach to recognise and classify speech events for people counting on embedded systems"

Symbolbild Informatik-Studierender mit Laptop in Sessel sitzend

Start date: immediately or by arrangement

Background:

The detection and analysis of acoustic events has become increasingly important in recent years thanks to advances in signal processing and machine learning. Particularly in the context of the Internet of Things (IoT), new embedded and networked devices open up innovative possibilities for monitoring and controlling the environment.

In the field of building technology, data protection-compliant methods are increasingly being used to determine the presence and number of people in rooms using unobtrusive audio sensors. Such information can be used, for example, to make the control of air conditioning systems more efficient and save energy.

As part of a research project with "Thermokon Sensortechnik GmbH", we are investigating the potential of audio-based methods for people counting. The aim is to use voice events as a basis for determining the number of people. The focus is on detecting, classifying and analysing voice signals in order to develop a reliable and resource-efficient system.

Aim of the master's thesis:

The aim of this master's thesis is to investigate and train machine learning models that are capable of reliably recognising and classifying speech events in audio signals. The aim is to create a robust basis for precise people counting. The following core tasks are involved in the work:

  1. Data preparation and analysis:
    • Collection and pre-processing of audio data, in particular speech events, taking into account ambient noise.
    • Investigation of suitable data formats
  2. Model training:
    • Comparison of modern neural network architectures (e.g. convolutional neural networks or recurrent neural networks) for the detection of speech events
    • Selection of one or more suitable models, e.g. Google YAMNet, taking into account the hardware requirements of embedded systems
    • Implementation and execution of the training process using a suitable machine learning library and a suitable data set.
  3. Classification:
    • Integration of a classifier that distinguishes speech events according to relevant features in order to enable the assignment and thus counting of persons.
  4. Evaluation:
    • Testing and validation of the developed models using real data sets.
    • Selection of suitable metrics in order to be able to make statements about accuracy, robustness and efficiency.
  5. Optimisation for embedded systems:
    • Application of techniques to optimise the models for resource-constrained devices.
    • Profiling of models on NPU-accelerated microcontrollers e.g. NXP, Rockchip

Prerequisites:

  • Good programming skills in Python, knowledge of C/C++ for embedded systems an advantage.
  • Experience in machine learning and interest in audio signal processing.
  • Experience with machine learning libraries (e.g. TensorFlow, PyTorch, Scikit-learn) is desirable.
  • Independent and structured way of working
  • Good written and spoken German and English language skills
  • Ability to work in a team and good communication skills

Additional information:

Prior preparation of an R&D study to familiarise yourself with the research question is also possible.

If you are interested, please contact Prof Dr Jens-Peter Akelbein.

Back