Tobias Bocklet

Sebastian P. Bayerl, Dominik Wagner, Elmar Nöth et al.

This paper empirically investigates the influence of different data splits and splitting strategies on the performance of dysfluency detection systems. For this, we perform experiments using wav2vec 2.0 models with a classification head as well as support vector machines (SVM) in conjunction with the features extracted from the wav2vec 2.0 model to detect dysfluencies. We train and evaluate the systems with different non-speaker-exclusive and speaker-exclusive splits of the Stuttering Events in Podcasts (SEP-28k) dataset to shed some light on the variability of results w.r.t. to the partition method used. Furthermore, we show that the SEP-28k dataset is dominated by only a few speakers, making it difficult to evaluate. To remedy this problem, we created SEP-28k-Extended (SEP-28k-E), containing semi-automatically generated speaker and gender information for the SEP-28k corpus, and suggest different data splits, each useful for evaluating other aspects of methods for dysfluency detection.

Sebastian P. Bayerl, Dominik Wagner, Ilja Baumann et al.

Vocal fatigue refers to the feeling of tiredness and weakness of voice due to extended utilization. This paper investigates the effectiveness of neural embeddings for the detection of vocal fatigue. We compare x-vectors, ECAPA-TDNN, and wav2vec 2.0 embeddings on a corpus of academic spoken English. Low-dimensional mappings of the data reveal that neural embeddings capture information about the change in vocal characteristics of a speaker during prolonged voice usage. We show that vocal fatigue can be reliably predicted using all three kinds of neural embeddings after only 50 minutes of continuous speaking when temporal smoothing and normalization are applied to the extracted embeddings. We employ support vector machines for classification and achieve accuracy scores of 81% using x-vectors, 85% using ECAPA-TDNN embeddings, and 82% using wav2vec 2.0 embeddings as input features. We obtain an accuracy score of 76%, when the trained system is applied to a different speaker and recording environment without any adaptation.

Franziska Braun, Sebastian P. Bayerl, Florian Hönig et al.

Speech pauses, alongside content and structure, offer a valuable and non-invasive biomarker for detecting dementia. This work investigates the use of pause-enriched transcripts in transformer-based language models to differentiate the cognitive states of subjects with no cognitive impairment, mild cognitive impairment, and Alzheimer's dementia based on their speech from a clinical assessment. We address three binary classification tasks: Onset, monitoring, and dementia exclusion. The performance is evaluated through experiments on a German Verbal Fluency Test and a Picture Description Test, comparing the model's effectiveness across different speech production contexts. Starting from a textual baseline, we investigate the effect of incorporation of pause information and acoustic context. We show the test should be chosen depending on the task, and similarly, lexical pause information and acoustic cross-attention contribute differently.

Ilja Baumann, Dominik Wagner, Sebastian Bayerl et al.

This work aims to automatically evaluate whether the language development of children is age-appropriate. Validated speech and language tests are used for this purpose to test the auditory memory. In this work, the task is to determine whether spoken nonwords have been uttered correctly. We compare different approaches that are motivated to model specific language structures: Low-level features (FFT), speaker embeddings (ECAPA-TDNN), grapheme-motivated embeddings (wav2vec 2.0), and phonetic embeddings in form of senones (ASR acoustic model). Each of the approaches provides input for VGG-like 5-layer CNN classifiers. We also examine the adaptation per nonword. The evaluation of the proposed systems was performed using recordings from different kindergartens of spoken nonwords. ECAPA-TDNN and low-level FFT features do not explicitly model phonetic information; wav2vec2.0 is trained on grapheme labels, our ASR acoustic model features contain (sub-)phonetic information. We found that the more granular the phonetic modeling is, the higher are the achieved recognition rates. The best system trained on ASR acoustic model features with VTLN achieved an accuracy of 89.4% and an area under the ROC (Receiver Operating Characteristic) curve (AUC) of 0.923. This corresponds to an improvement in accuracy of 20.2% and AUC of 0.309 relative compared to the FFT-baseline.

Christopher Simic, Korbinian Riedhammer, Tobias Bocklet

The FAME 2026 challenge comprises two demanding tasks: training face-voice associations combined with a multilingual setting that includes testing on languages on which the model was not trained. Our approach consists of separate uni-modal processing pipelines with general face and voice feature extraction, complemented by additional age-gender feature extraction to support prediction. The resulting single-modal features are projected into a shared embedding space and trained with an Adaptive Angular Margin (AAM) loss. Our approach achieved first place in the FAME 2026 challenge, with an average Equal-Error Rate (EER) of 23.99%.

Steffen Freisinger, Philipp Seeberger, Thomas Ranzenberger et al.

Segmenting speech transcripts into thematic sections benefits both downstream processing and users who depend on written text for accessibility. We introduce a novel approach to hierarchical topic segmentation in transcripts, generating multi-level tables of contents that capture both topic and subtopic boundaries. We compare zero-shot prompting and LoRA fine-tuning on large language models, while also exploring the integration of high-level speech pause features. Evaluations on English meeting recordings and multilingual lecture transcripts (Portuguese, German) show significant improvements over established topic segmentation baselines. Additionally, we adapt a common evaluation measure for multi-level segmentation, taking into account all hierarchical levels within one metric.

Steffen Freisinger, Philipp Seeberger, Tobias Bocklet et al.

Spoken content, such as online videos and podcasts, often spans multiple topics, which makes automatic topic segmentation essential for user navigation and downstream applications. However, current methods do not fully leverage acoustic features, leaving room for improvement. We propose a multi-modal approach that fine-tunes both a text encoder and a Siamese audio encoder, capturing acoustic cues around sentence boundaries. Experiments on a large-scale dataset of YouTube videos show substantial gains over text-only and multi-modal baselines. Our model also proves more resilient to ASR noise and outperforms a larger text-only baseline on three additional datasets in Portuguese, German, and English, underscoring the value of learned acoustic features for robust topic segmentation.

Philipp Seeberger, Steffen Freisinger, Tobias Bocklet et al.

Due to the rapid growth of social media platforms, these tools have become essential for monitoring information during ongoing disaster events. However, extracting valuable insights requires real-time processing of vast amounts of data. A major challenge in existing systems is their exposure to event-related biases, which negatively affects their ability to generalize to emerging events. While recent advancements in debiasing and causal learning offer promising solutions, they remain underexplored in the disaster event domain. In this work, we approach bias mitigation through a causal lens and propose a method to reduce event- and domain-related biases, enhancing generalization to future events. Our approach outperforms multiple baselines by up to +1.9% F1 and significantly improves a PLM-based classifier across three disaster classification tasks.

Dong Bok Lee, Seanie Lee, Sangwoo Park et al.

The reliability of large language models (LLMs) during test-time scaling is often assessed with \emph{external verifiers} or \emph{reward models} that distinguish correct reasoning from flawed logic. Prior work generally assumes that process reward models (PRMs), which score every intermediate reasoning step, outperform outcome reward models (ORMs) that assess only the final answer. This view is based mainly on evidence from narrow, math-adjacent domains. We present the first unified evaluation of four reward model variants, discriminative ORM and PRM (\DisORM, \DisPRM) and generative ORM and PRM (\GenORM, \GenPRM), across 14 diverse domains. Contrary to conventional wisdom, we find that (i) \DisORM performs on par with \DisPRM, (ii) \GenPRM is not competitive, and (iii) overall, \GenORM is the most robust, yielding significant and consistent gains across every tested domain. We attribute this to PRM-style stepwise scoring, which inherits label noise from LLM auto-labeling and has difficulty evaluating long reasoning trajectories, including those involving self-correcting reasoning. Our theoretical analysis shows that step-wise aggregation compounds errors as reasoning length grows, and our empirical observations confirm this effect. These findings challenge the prevailing assumption that fine-grained supervision is always better and support generative outcome verification for multi-domain deployment. We publicly release our code, datasets, and checkpoints at \href{https://github.com/db-Lee/Multi-RM}{\underline{\small\texttt{https://github.com/db-Lee/Multi-RM}}} to facilitate future research in multi-domain settings.

Roland Kammerbauer, Thomas H. Schmitt, Tobias Bocklet

In the woodworking industry, a huge amount of effort has to be invested into the initial quality assessment of the raw material. In this study we present an AI model to detect, quantify and localize defects on wooden logs. This model aims to both automate the quality control process and provide a more consistent and reliable quality assessment. For this purpose a dataset of 1424 sample images of wood logs is created. A total of 5 annotators possessing different levels of expertise is involved in dataset creation. An inter-annotator agreement analysis is conducted to analyze the impact of expertise on the annotation task and to highlight subjective differences in annotator judgement. We explore, train and fine-tune the state-of-the-art InternImage and ONE-PEACE architectures for semantic segmentation. The best model created achieves an average IoU of 0.71, and shows detection and quantification capabilities close to the human annotators.

Maximilian Bundscherer, Thomas H. Schmitt, Tobias Bocklet

In industrial manufacturing of glass bottles, quality control of bottle prints is necessary as numerous factors can negatively affect the printing process. Even minor defects in the bottle prints must be detected despite reflections in the glass or manufacturing-related deviations. In cooperation with our medium-sized industrial partner, two ML-based approaches for quality control of these bottle prints were developed and evaluated, which can also be used in this challenging scenario. Our first approach utilized different filters to supress reflections (e.g. Sobel or Canny) and image quality metrics for image comparison (e.g. MSE or SSIM) as features for different supervised classification models (e.g. SVM or k-Neighbors), which resulted in an accuracy of 84%. The images were aligned based on the ORB algorithm, which allowed us to estimate the rotations of the prints, which may serve as an indicator for anomalies in the manufacturing process. In our second approach, we fine-tuned different pre-trained CNN models (e.g. ResNet or VGG) for binary classification, which resulted in an accuracy of 87%. Utilizing Grad-Cam on our fine-tuned ResNet-34, we were able to localize and visualize frequently defective bottle print regions. This method allowed us to provide insights that could be used to optimize the actual manufacturing process. This paper also describes our general approach and the challenges we encountered in practice with data collection during ongoing production, unsupervised preselection, and labeling.

Thomas H. Schmitt, Maximilian Bundscherer, Tobias Bocklet

Monitoring leftover products provides valuable insights that can be used to optimize future production. This is especially important for German bakeries because freshly baked goods have a very short shelf life. Automating this process can reduce labor costs, improve accuracy, and streamline operations. We propose automating this process using an object detection model to identify baked goods from images. However, the large diversity of German baked goods makes fully supervised training prohibitively expensive and limits scalability. Although open-vocabulary detectors (e.g., OWLv2, Grounding DINO) offer lexibility, we demonstrate that they are insufficient for our task. While motivated by bakeries, our work addresses the broader challenges of deploying computer vision in industries, where tasks are specialized and annotated datasets are scarce. We compile dataset splits with varying supervision levels, covering 19 classes of baked goods. We propose two training workflows to train an object detection model with limited supervision. First, we combine OWLv2 and Grounding DINO localization with image-level supervision to train the model in a weakly supervised manner. Second, we improve viewpoint robustness by fine-tuning on video frames annotated using Segment Anything 2 as a pseudo-label propagation model. Using these workflows, we train YOLOv11 for our detection task due to its favorable speed accuracy tradeoff. Relying solely on image-level supervision, the model achieves a mean Average Precision (mAP) of 0.91. Finetuning with pseudo-labels raises model performance by 19.3% under non-ideal deployment conditions. Combining these workflows trains a model that surpasses our fully-supervised baseline model under non-ideal deployment conditions, despite relying only on image-level supervision.

Thomas H. Schmitt, Maximilian Bundscherer, Tobias Bocklet

In the food industry, reprocessing returned product is a vital step to increase resource efficiency. [SBB23] presented an AI application that automates the tracking of returned bread buns. We extend their work by creating an expanded dataset comprising 2432 images and a wider range of baked goods. To increase model robustness, we use generative models pix2pix and CycleGAN to create synthetic images. We train state-of-the-art object detection model YOLOv9 and YOLOv8 on our detection task. Our overall best-performing model achieved an average precision AP@0.5 of 90.3% on our test set.

Natalie Engert, Dominik Wagner, Korbinian Riedhammer et al.

Wav2vec 2.0 (W2V2) has shown strong performance in pathological speech analysis by effectively capturing the characteristics of atypical speech. Despite its success, it remains unclear which components of its learned representations are most informative for specific downstream tasks. In this study, we address this question by investigating the regression of dysarthric speech descriptors using annotations from the Speech Accessibility Project dataset. We focus on five descriptors, each addressing a different aspect of speech or voice production: intelligibility, imprecise consonants, inappropriate silences, harsh voice and monoloudness. Speech representations are derived from a W2V2-based feature extractor, and we systematically compare layer-wise and time-wise aggregation strategies using attentive statistics pooling. Our results show that intelligibility is best captured through layer-wise representations, whereas imprecise consonants, harsh voice and monoloudness benefit from time-wise modeling. For inappropriate silences, no clear advantage could be observed for either approach.

Seanie Lee, Dong Bok Lee, Dominik Wagner et al.

Deploying large language models (LLMs) in real-world applications requires robust safety guard models to detect and block harmful user prompts. While large safety guard models achieve strong performance, their computational cost is substantial. To mitigate this, smaller distilled models are used, but they often underperform on "hard" examples where the larger model provides accurate predictions. We observe that many inputs can be reliably handled by the smaller model, while only a small fraction require the larger model's capacity. Motivated by this, we propose SafeRoute, a binary router that distinguishes hard examples from easy ones. Our method selectively applies the larger safety guard model to the data that the router considers hard, improving efficiency while maintaining accuracy compared to solely using the larger safety guard model. Experimental results on multiple benchmark datasets demonstrate that our adaptive model selection significantly enhances the trade-off between computational cost and safety performance, outperforming relevant baselines.

Seanie Lee, Sangwoo Park, Dong Bok Lee et al.

Low-Rank Adaptation (LoRA), which introduces a product of two trainable low-rank matrices into frozen pre-trained weights, is widely used for efficient fine-tuning of language models in federated learning (FL). However, when combined with differentially private stochastic gradient descent (DP-SGD), LoRA faces substantial noise amplification: DP-SGD perturbs per-sample gradients, and the matrix multiplication of the LoRA update ($BA$) intensifies this effect. Freezing one matrix (e.g., $A$) reduces the noise but restricts model expressiveness, often resulting in suboptimal adaptation. To address this, we propose $\texttt{FedSVD}$, a simple yet effective method that introduces a global reparameterization based on singular value decomposition (SVD). In our approach, each client optimizes only the $B$ matrix and transmits it to the server. The server aggregates the $B$ matrices, computes the product $BA$ using the previous $A$, and refactorizes the result via SVD. This yields a new adaptive $A$ composed of the orthonormal right singular vectors of $BA$, and an updated $B$ containing the remaining SVD components. This reparameterization avoids quadratic noise amplification, while allowing $A$ to better capture the principal directions of the aggregate updates. Moreover, the orthonormal structure of $A$ bounds the gradient norms of $B$ and preserves more signal under DP-SGD, as confirmed by our theoretical analysis. As a result, $\texttt{FedSVD}$ consistently improves stability and performance across a variety of privacy settings and benchmarks, outperforming relevant baselines under both private and non-private regimes.

Christopher Simic, Korbinian Riedhammer, Tobias Bocklet

We present an approach to Audio-Visual Speech Recognition that builds on a pre-trained Whisper model. To infuse visual information into this audio-only model, we extend it with an AV fusion module and LoRa adapters, one of the most up-to-date adapter approaches. One advantage of adapter-based approaches, is that only a relatively small number of parameters are trained, while the basic model remains unchanged. Common AVSR approaches train single models to handle several noise categories and noise levels simultaneously. Taking advantage of the lightweight nature of adapter approaches, we train noise-scenario-specific adapter-sets, each covering individual noise-categories or a specific noise-level range. The most suitable adapter-set is selected by previously classifying the noise-scenario. This enables our models to achieve an optimum coverage across different noise-categories and noise-levels, while training only a minimum number of parameters. Compared to a full fine-tuning approach with SOTA performance our models achieve almost comparable results over the majority of the tested noise-categories and noise-levels, with up to 88.5% less trainable parameters. Our approach can be extended by further noise-specific adapter-sets to cover additional noise scenarios. It is also possible to utilize the underlying powerful ASR model when no visual information is available, as it remains unchanged.

Philipp Seeberger, Tobias Bocklet, Korbinian Riedhammer

User-generated information content has become an important information source in crisis situations. However, classification models suffer from noise and event-related biases which still poses a challenging task and requires sophisticated task-adaptation. To address these challenges, we propose the use of contrastive task-specialized sentence encoders for downstream classification. We apply the task-specialization on the CrisisLex, HumAID, and TrecIS information type classification tasks and show performance gains w.r.t. F1-score. Furthermore, we analyse the cross-corpus and cross-lingual capabilities for two German event relevancy classification datasets.

Maximilian Bundscherer, Thomas H. Schmitt, Ilja Baumann et al.

This study presents an ML approach for classifying digital radio operating modes evaluated on real-world transmissions. We generated 98 different parameterized radio signals from 17 digital operating modes, transmitted each of them on the 70 cm (UHF) amateur radio band, and recorded our transmissions with two different architectures of SDR receivers. Three lightweight ML models were trained exclusively on spectrograms of limited non-transmitted signals with random characters as payloads. This training involved an online data augmentation pipeline to simulate various radio channel impairments. Our best model, EfficientNetB0, achieved an accuracy of 93.80% across the 17 operating modes and 85.47% across all 98 parameterized radio signals, evaluated on our real-world transmissions with Wikipedia articles as payloads. Furthermore, we analyzed the impact of varying signal durations & the number of FFT bins on classification, assessed the effectiveness of our simulated channel impairments, and tested our models across multiple simulated SNRs.

Dominik Wagner, Sebastian P. Bayerl, Ilja Baumann et al.

Accurately detecting dysfluencies in spoken language can help to improve the performance of automatic speech and language processing components and support the development of more inclusive speech and language technologies. Inspired by the recent trend towards the deployment of large language models (LLMs) as universal learners and processors of non-lexical inputs, such as audio and video, we approach the task of multi-label dysfluency detection as a language modeling problem. We present hypotheses candidates generated with an automatic speech recognition system and acoustic representations extracted from an audio encoder model to an LLM, and finetune the system to predict dysfluency labels on three datasets containing English and German stuttered speech. The experimental results show that our system effectively combines acoustic and lexical information and achieves competitive results on the multi-label stuttering detection task.

Dominik Wagner, Seanie Lee, Ilja Baumann et al.

In this work, we optimize speculative sampling for parallel hardware accelerators to improve sampling speed. We notice that substantial portions of the intermediate matrices necessary for speculative sampling can be computed concurrently. This allows us to distribute the workload across multiple GPU threads, enabling simultaneous operations on matrix segments within thread blocks. This results in profiling time improvements ranging from 6% to 13% relative to the baseline implementation, without compromising accuracy. To further accelerate speculative sampling, probability distributions parameterized by softmax are approximated by sigmoid. This approximation approach results in significantly greater relative improvements in profiling time, ranging from 37% to 94%, with a minor decline in accuracy. We conduct extensive experiments on both automatic speech recognition and summarization tasks to validate the effectiveness of our optimization methods.

Thomas H. Schmitt, Maximilian Bundscherer, Tobias Bocklet

The Semmeldetector, is a machine learning application that utilizes object detection models to detect, classify and count baked goods in images. Our application allows commercial bakers to track unsold baked goods, which allows them to optimize production and increase resource efficiency. We compiled a dataset comprising 1151 images that distinguishes between 18 different types of baked goods to train our detection models. To facilitate model training, we used a Copy-Paste augmentation pipeline to expand our dataset. We trained the state-of-the-art object detection model YOLOv8 on our detection task. We tested the impact of different training data, model scale, and online image augmentation pipelines on model performance. Our overall best performing model, achieved an AP@0.5 of 89.1% on our test set. Based on our results, we conclude that machine learning can be a valuable tool even for unforeseen industries like bakeries, even with very limited datasets.

Sebastian P. Bayerl, Dominik Wagner, Ilja Baumann et al.

Most stuttering detection and classification research has viewed stuttering as a multi-class classification problem or a binary detection task for each dysfluency type; however, this does not match the nature of stuttering, in which one dysfluency seldom comes alone but rather co-occurs with others. This paper explores multi-language and cross-corpus end-to-end stuttering detection as a multi-label problem using a modified wav2vec 2.0 system with an attention-based classification head and multi-task learning. We evaluate the method using combinations of three datasets containing English and German stuttered speech, one containing speech modified by fluency shaping. The experimental results and an error analysis show that multi-label stuttering detection systems trained on cross-corpus and multi-language data achieve competitive results but performance on samples with multiple labels stays below over-all detection results.

Munir Georges, Jonathan Huang, Tobias Bocklet

Deep neural networks (DNN) have recently been widely used in speaker recognition systems, achieving state-of-the-art performance on various benchmarks. The x-vector architecture is especially popular in this research community, due to its excellent performance and manageable computational complexity. In this paper, we present the lrx-vector system, which is the low-rank factorized version of the x-vector embedding network. The primary objective of this topology is to further reduce the memory requirement of the speaker recognition system. We discuss the deployment of knowledge distillation for training the lrx-vector system and compare against low-rank factorization with SVD. On the VOiCES 2019 far-field corpus we were able to reduce the weights by 28% compared to the full-rank x-vector system while keeping the recognition rate constant (1.83% EER).