Themen
Auf dieser Seite finden Sie aktuelle Themen für Abschlussarbeiten (Bachelor, Master). Sollten Sie eigene Ideen für Themen im Umfeld der Forschungsthemen der Professur Computational Humanities haben, kontaktieren Sie uns gerne.
Topic:Automatic Melody Simplification for Music Learning
Description:
Learning to play melodic material can be challenging due to wide pitch ranges, large intervallic leaps, rhythmic complexity, or technically demanding note sequences. Automatic melody simplification aims to transform an original melody into an easier-to-play version while preserving its musical identity. This master’s thesis investigates state-of-the-art methods for melody simplification in a music educational context, with the goal of making melodies more accessible for learners. The work will be carried out in close cooperation with the Semantic Media Technologies (SMT) group at the Fraunhofer Institute for Digital Media Technologies (IDMT), including a 1–2 month research visit and co-supervision by Dr. Andrew McLeod. The thesis will implement one to two baseline systems from the literature (e.g., rule-based or data-driven approaches) and extend them with novel constraints or learning objectives tailored to educational use cases. Different simplified melody variants will be systematically evaluated against the original using perceptual listening experiments, such as a MUSHRA-style test or triplet-based similarity annotations, to assess musical similarity and perceived learning difficulty.
Keywords: melody simplification, music education, music information retrieval (MIR), perceptual evaluation, listening tests
---
Topic: Automatic String Tuning Estimation for Guitar Recordings
Description: Guitar performances involve multiple strings that can be tuned in various ways, creating challenges for automatic analysis and transcription. Since the same pitch can often be played at different fretboard positions, distinguishing between standard and open tunings, which are frequently used to simplify chord fingering, remains an open problem in music information retrieval (MIR). This thesis investigates methods for estimating the tuning configuration of string instruments directly from audio recordings, building upon existing datasets such as the IDMT Guitar Dataset and recordings by the Peruvian guitarist Daniel Kirwayo. The work may combine data-driven audio analysis and symbolic approaches to explore how tuning affects fingering effort and chord formation.
Keywords: tuning estimation, guitar analysis, transcription, music information retrieval (MIR)
---
Topic: Automatic Guitar Tablature Generation
Description: Translating musical notes into guitar tablature is a challenging task due to the ambiguity between pitch and playing position—the same note can be produced on multiple strings and frets. Guitarists typically select fingerings that minimize finger stretching and movement across the fretboard, making tablature generation a complex optimization problem constrained by tuning, pitch range, and playability. This thesis investigates methods for automatically generating guitar tablature from symbolic music scores or outputs of automatic music transcription systems, aiming to produce realistic and ergonomic fingerings that reflect how musicians actually play.
Keywords: guitar tablature, automatic music transcription, fingering optimization, music information retrieval (MIR), performance modeling
---
Topic: Linking Perception and Recognition — Psychoacoustic Cues in Urban Soundscapes
Description: How do people’s perceptual impressions of urban soundscapes relate to what machines detect in the same recordings? This thesis explores the relationship between psychoacoustic parameters (such as loudness, sharpness, roughness, and tonality) and sound event categories identified by state-of-the-art deep learning models (e.g., PANN, CLAP). Using publicly available datasets like TAU Urban Acoustic Scenes 2019 or SONYC Urban Sound Tagging (SONYC-UST), the work aims to reveal typical co-occurrence patterns between perceptual attributes, sound classes, and acoustic scenes to deepen our understanding of urban acoustic environments.
Keywords: soundscape, psychoacoustics, sound event detection, DCASE
----
Topic: Soundscapes of Bamberg — Mapping Urban Acoustic Environments
Description: How does the sound of a city change throughout the day, and how do its residents perceive these variations? This thesis explores the urban soundscapes of Bamberg by combining geospatial data, citizen surveys, and immersive audio recordings. Based on an initial survey and geodata analysis, characteristic locations with distinct time-of-day sound compositions will be identified and recorded using binaural or ambisonic microphones. These soundscapes will then be made explorable in an interactive web interface, accompanied by visualizations of survey results that reflect the affective impact of different sounds and environments on residents.
Keywords: soundscape analysis, urban acoustics, geodata, visualization
----
Topic: Synthesis and Evaluation of Natural Sounds in Urban Soundscapes
Description: This thesis will investigate the synthesis and reconstruction of natural sounds in urban soundscapes, which consist of various static and moving sources, from environmental noise to human activity. In line with ecological soundscape research, sounds will be considered as geophonic (non-biological natural sounds, e.g., rainfall, wind), biophonic (produced by non-human organisms), and anthrophonic (human-generated sounds, e.g., traffic, machinery). The thesis will include a literature review of state-of-the-art synthesis methods, such as those developed for the Foley task in the DCASE Challenge 2023. It will define a meaningful categorization of urban sound types and collect a dataset of exemplary sounds, for instance from FSD50k. Finally, the thesis will evaluate existing synthesis methods and conduct a listening test to assess the naturalness, realism, and clarity of the generated sounds.
Keywords: Urban soundscapes, Sound synthesis, Perceptual evaluation
----
Topic: Automatic Detection of Pronunciation Errors in L2 English
This project will be co-supervised by Prof. Dr Julia Schlüter, Chair of English Linguistics at the University of Bamberg.
Description: This thesis will investigate how machine learning can automatically identify pronunciation errors made by advanced German learners of English. It is based on a novel dataset consisting of 250 one-minute recordings of a fixed diagnostic text. These recordings are annotated with 34 categories of typical learner errors and are accompanied by an IPA transcription. The thesis will involve aligning spectrogram representations of the recordings with the IPA reference text, extracting relevant acoustic–phonetic features and training a model to recognise mispronunciations such as final devoicing, voicing confusions, /v/–/w/ substitutions or vowel mergers. An alternative approach would be to investigate whether a speech recognition model such as Whisper could be fine-tuned for this task. The project also aims to evaluate the accuracy of detection across error types and analyse which acoustic cues most reliably distinguish correct from incorrect pronunciations. This will provide insights relevant to automatic pronunciation assessment.
Keywords: pronunciation error detection, L2 English, IPA alignment, acoustic modelling, speech analysis, Whisper
----
Topic: Predictive Maintenance for Industrial Manufacturing Processes Using Audio and Multimodal Analysis
Description: Modern industrial production systems require reliable monitoring to detect faults early and reduce unplanned downtime. Predictive maintenance aims to identify anomalies and degradation patterns before failures occur, but real-world deployments are challenged by limited labeled data, domain shifts across machines and processes, and complex sensor conditions. This master’s thesis investigates machine learning methods for predictive maintenance in industrial settings, with a focus on acoustic and multimodal sensing. Possible subtopics include acoustic anomaly detection, few-shot learning for rare fault conditions, domain adaptation and generalization, and multimodal learning that combines audio and visual information. The work targets representative manufacturing processes such as Schweißen (MIG/MAG welding, laser beam welding) and Zerspanen (milling and turning). The thesis will develop a prototypical end-to-end system based on state-of-the-art methods and evaluate it under realistic conditions using real-world industrial data. The work will be co-supervised by Dr.-Ing. Sascha Grollmisch from the Industrial Media Applications group at Fraunhofer IDMT, with a strong emphasis on practical relevance and empirical validation.
Keywords: predictive maintenance, acoustic anomaly detection, industrial monitoring, few-shot learning, multimodal machine learning