Topics for theses and student projects
How can data streams remain reliable under changing network conditions? How does a system recover the right information from a conversation and distinguish its speakers? These topics connect transport protocols and edge computing with speech, text and audio processing. Student projects also offer a practical route through measurement tools, sensors and reproducible experiments.
Bachelor’s theses focus on understandable implementations, comparisons and error analysis. Master’s theses develop and investigate independent methods or substantive adaptations. A student project focuses on a working component with understandable tests. Scope, data and supervision are agreed in discussion to suit your degree programme and available time.
Bachelor’s theses
Network congestion control
How do conventional congestion-control methods perform when traffic conditions change? This thesis compares their behaviour through simulation and performance evaluation.
Tasks
- Review existing congestion-control approaches.
- Implement conventional algorithms in a simulation environment.
- Compare their performance under different network conditions.
Requirements: Basic knowledge of Python, Linux and Git.
Assistant: Dr. Ali Ghermezian
Research context: RMDT · AI-RMDT
Task offloading in edge computing
Which tasks should run at the edge, and which should be sent to the cloud? This thesis investigates task-offloading strategies by evaluating task execution and resource use across edge and cloud environments.
Tasks
- Review existing task-offloading strategies.
- Implement basic offloading methods using simulation tools.
- Evaluate latency and resource utilisation.
- Compare the results under different workloads.
Requirements: Basic Python programming and an interest in distributed systems.
Research context: Networked systems and intelligent infrastructure
Impact of Speech Recognition Errors on Information Extraction from Dialogues
A spoken place name can become a different word in a transcript. When does that error also change the travel information a system extracts from the conversation? This bachelor’s thesis investigates how speech-recognition errors affect departure places, destinations, departure times and arrival deadlines in English taxi and train dialogues—and whether earlier context helps or carries previous mistakes forward.
Tasks
- Prepare a manageable SpokenWOZ dialogue subset; check reference transcripts by listening and verify the expected field values independently.
- Use one pretrained ASR system and one fixed extractor. Compare checked and automatically recognised text on the same turns, each with and without the preceding dialogue history.
- Evaluate word error rate, field-value precision, recall and F1, and exact match of all four fields. Analyse places, times, negation and corrections through paired error examples.
- Deliver code, configurations, predictions and a reproducible evaluation report, keeping development and test dialogues separate.
The contribution is a controlled explanation of error propagation. Components stay fixed; training an ASR model or inventing a new extraction method is outside the required scope.
Assistant: Anastasiia Sapeha
Research context: AI and multimodal analysis
Robustness of Travel Information Extraction from Text Dialogues
“From the station to the museum, no later than six” contains places and a time constraint. A conversation adds references, paraphrases and later corrections. This bachelor’s thesis compares how reliably different methods recover that information from English text dialogues, and how available context affects both accuracy and runtime.
Tasks
- Predict the same four fields for taxi and train dialogues in MultiWOZ 2.2: departure, destination, departure time and arrival deadline. Use only the history visible at each evaluated turn.
- Compare three complete methods from distinct families, including a rule-based baseline. Give every method the same input and output schema.
- Study checked paraphrases, consistent place substitutions and indirect references. Compare the current utterance, a short context and the full visible dialogue prefix.
- Evaluate field values with precision, recall, F1 and exact match; record runtime, failures and unsupported outputs. Deliver reproducible code and a paired error analysis.
The expected contribution is a defensible comparison of robustness and computational cost. A new large model is not required; fine-tuning and a second corpus are optional extensions.
Assistant: Anastasiia Sapeha
Research context: AI and multimodal analysis
Master’s theses
Reinforcement learning for network congestion control
Can reinforcement learning adapt congestion control to changing traffic conditions? This thesis develops and evaluates an adaptive algorithm in simulation, comparing it with conventional methods.
Tasks
- Review existing congestion-control approaches.
- Develop and implement a reinforcement-learning algorithm in a simulation environment.
- Optimise network performance under dynamic traffic conditions.
- Compare the results with conventional methods.
Requirements: Knowledge of reinforcement learning and working knowledge of Python, Linux and Git.
Assistant: Dr. Ali Ghermezian
Research context: AI-RMDT · AI and multimodal analysis
Reinforcement learning for edge computing
Can a learning-based decision improve where a task is executed? This thesis investigates reinforcement learning for task offloading, with the aim of improving task execution and resource utilisation across edge and cloud environments.
Tasks
- Review existing task-offloading strategies.
- Develop a reinforcement-learning approach using simulation tools.
- Compare the results with existing methods.
Requirements: Knowledge of reinforcement learning and working knowledge of Python, Linux and Git.
Research context: Networked systems and intelligent infrastructure · AI and multimodal analysis
Removing Background Music while Preserving Dialogue
Two people are talking while a song plays in the background. How can the whole song, including the singer, be removed without damaging quiet voices or overlapping turns? This master’s thesis combines song identification, reference-based cancellation and dialogue separation with an independently designed strategy for protecting conversational speech.
Tasks
- Build a local catalogue of freely usable music; identify the song and matching excerpt, then adapt its reference to the recording’s delay, level and acoustic path.
- Establish adaptive cancellation and a pretrained dialogue-separation model as comparisons. Include the unprocessed mixture and simple aligned subtraction.
- Develop a confidence or adaptation strategy that protects dialogue when a reference is wrong or the acoustic path changes. Test its contribution through controlled ablations.
- Evaluate controlled conversation–music mixtures, for example using AMI and MUSAN. Examine remaining instruments, singing, speech damage and overlap separately, using SI-SDR, supplementary STOI on single-speaker intervals and listening tests.
The outcome is an executable pipeline and a justified methodological study. The required scope is offline and single-channel; real-time operation and training a new separator are optional. Better suppression is a hypothesis to test.
Assistant: Anastasiia Sapeha
Research context: AI and multimodal analysis
Text-Based Speaker Diarization and Contextual Name Attribution
Who is speaking when only the words of a conversation are available? This master’s thesis develops a text-only method that jointly estimates the number of active speakers and consistent utterance assignments. Names are linked afterwards, and only when the supplied dialogue provides evidence for them.
Tasks
- Prepare English conversation excerpts, starting with AMI, using neutral utterance IDs. Exclude audio, the true speaker count, roles and other metadata from model input.
- Compare text-embedding clustering and direct instruction-model inference as two baseline families.
- Develop an independent method or substantive adaptation for consistency across the whole excerpt, for example using pairwise speaker scores or revisable speaker memory. Investigate individual components through ablations.
- Evaluate speaker count, grouping and name attribution separately. Analyse merges, unnecessary splits, returning participants and unsupported names; support each assigned name with specific text evidence.
The expected contribution is an independent methodological study with reproducible inference and evaluation code. Unsupported names stay unknown. Training is allowed but not mandatory; a software demonstration alone does not replace the investigation.
Assistant: Anastasiia Sapeha
Research context: AI and multimodal analysis
Student projects
Making network experiments repeatable
Student project. Build an executable workflow for defined bandwidth, delay and loss profiles. Each run should keep its configuration, software version and measurement files together and reproduce the analysis steps from the same inputs. A small set of documented test cases is more useful than a large interface without a verifiable workflow. This topic suits students with some scripting experience.
Research context: RMDT · AI-RMDT · LTest · HighPerTimer · Experimental facilities
From a sensor node to a traceable measurement log
Student project. Connect a small sensor experiment to a receiver and a structured log. Detect missing messages, record transmission settings and show the relationship between the measured value, transmission and reception. The task can begin with simulated data and move to available devices by agreement. Initial experience with microcontrollers or Python is enough for a bounded starting task.
Research context: Networked systems and intelligent infrastructure · SmartLighting · Experimental facilities
A small test dataset with two sources
Student project. Create a manageable dataset with two related sources, such as synthetic movement sequences and short speech cues. Define timing, file formats and test cases, then build an analysis that reveals missing or contradictory associations. The aim is a well-documented component for later experiments. Programming experience and familiarity with data formats are helpful.
Research context: AI and multimodal analysis · AURORA · ReSeDiUm
A Reproducible Benchmark for Open-Set Speaker Identification
A system should recognise an enrolled voice and also reject someone it has never enrolled. This student project builds a reusable benchmark for that decision. Supplied single-speaker segments are compared with a fixed gallery of voice profiles to study the effects of enrollment duration, probe length and noise.
Tasks
- Build voice profiles using pretrained encoders. Compare ECAPA-TDNN with one alternative under identical preprocessing and trials, keeping their weights fixed.
- Create a reproducible split, starting with LibriSpeech and MUSAN noise. Keep validation and test speakers disjoint and separate enrollment recordings from probes.
- Choose one threshold per model on validation data, then freeze it for all test conditions. The calibration target of at most five percent unknown false acceptance is not a guarantee for the test rate.
- Report correct identification, rejection of known speakers, acceptance of unknown speakers and wrong accepted identities separately. Cache scores and automatically report metrics, plots, runtime and failures.
The result is tested code, configurations, split manifests and an automatically reproducible report. Identification and rejection use supplied segments; automatic segmentation and new model training are optional extensions.
Assistant: Anastasiia Sapeha
Research context: AI and multimodal analysis
Reproducing Speaker Attribution on Fisher and AMI
Can the words alone reveal whether two text samples come from the same person? This student project reproduces the TACL study “Can Authorship Attribution Models Distinguish Speakers in Speech Transcripts?” on Fisher telephone conversations, then applies the same methods to AMI meetings. The text samples are already grouped by their true speakers: the task verifies two supplied samples.
Tasks
- Reproduce the study’s six methods—TF-IDF, PANgrams, SBERT, CISR, LUAR and AdHominem—and six experiment blocks. Check model availability and compute needs in a preliminary run.
- Reconstruct the original speaker splits, trial construction and settings. Investigate transcript style, topic control, fine-tuning, pretraining domain and input size as specified in the paper.
- Construct AMI pairs across meetings and with within-meeting negatives. Keep recurring participants and connected teams within one split, and state where AMI conditions differ from Fisher.
- Evaluate without AMI tuning first. Report ROC AUC, EER and errors at a development-selected fixed threshold; regenerate tables and figures and explain discrepancies or unavailable results.
The outcome is an executable reproduction and a documented corpus comparison. Speaker turns are not estimated here. Licensed Fisher transcripts remain in authorised storage; a new attribution method or higher accuracy is not required.
Assistant: Anastasiia Sapeha
Research context: AI and multimodal analysis
Discuss a topic
To discuss a topic, contact Prof. Dr. Dmitry Kachan first: Dmitry.Kachan@hs-anhalt.de.
When enquiring, include your degree programme, the kind of work you are considering, your interests and experience, and your proposed timeframe. If you already have a question or a suitable dataset, briefly describe both. The contacts in the FILA team can help turn them into a workable task.
