Paula Andrea Pérez-Toro

Abner Hernandez, Paula Andrea Pérez-Toro, Elmar Nöth et al.

State-of-the-art automatic speech recognition (ASR) systems perform well on healthy speech. However, the performance on impaired speech still remains an issue. The current study explores the usefulness of using Wav2Vec self-supervised speech representations as features for training an ASR system for dysarthric speech. Dysarthric speech recognition is particularly difficult as several aspects of speech such as articulation, prosody and phonation can be impaired. Specifically, we train an acoustic model with features extracted from Wav2Vec, Hubert, and the cross-lingual XLSR model. Results suggest that speech representations pretrained on large unlabelled data can improve word error rate (WER) performance. In particular, features from the multilingual model led to lower WERs than filterbanks (Fbank) or models trained on a single language. Improvements were observed in English speakers with cerebral palsy caused dysarthria (UASpeech corpus), Spanish speakers with Parkinsonian dysarthria (PC-GITA corpus) and Italian speakers with paralysis-based dysarthria (EasyCall corpus). Compared to using Fbank features, XLSR-based features reduced WERs by 6.8%, 22.0%, and 7.0% for the UASpeech, PC-GITA, and EasyCall corpus, respectively.

Serli Kopar, Roshan Prakash Rane, Christian Mychajliw et al.

This study examines the relationship between speech representations and the hierarchical structure of cognitive assessment in mild cognitive impairment. Utilizing 5,754 German neuropsychological assessment recordings, we evaluate six cognitive tasks across three score levels: task, domain, and global levels. We compare hand-crafted acoustic features with self-supervised learning (SSL) embeddings. Results show that although SSL representations generally outperform hand-crafted features at lower levels, this trend reverses for MCI classification. Furthermore, task-specific constraints influence performance: tasks with greater response freedom exhibit performance dilution as hierarchical levels increase, suggesting ``specialist'' representations, whereas the performance of highly structured tasks increases toward higher levels, suggesting ``generalist'' representations. These findings show links between task constraints and assessment hierarchy in automated clinical speech analysis.

Soroosh Tayebi Arasteh, Mahshad Lotfinia, Paula Andrea Perez-Toro et al.

Speech pathology has impacts on communication abilities and quality of life. While deep learning-based models have shown potential in diagnosing these disorders, the use of sensitive data raises critical privacy concerns. Although differential privacy (DP) has been explored in the medical imaging domain, its application in pathological speech analysis remains largely unexplored despite the equally critical privacy concerns. To the best of our knowledge, this study is the first to investigate DP's impact on pathological speech data, focusing on the trade-offs between privacy, diagnostic accuracy, and fairness. Using a large, real-world dataset of 200 hours of recordings from 2,839 German-speaking participants, we observed a maximum accuracy reduction of 3.85% when training with DP with high privacy levels. To highlight real-world privacy risks, we demonstrated the vulnerability of non-private models to gradient inversion attacks, reconstructing identifiable speech samples and showcasing DP's effectiveness in mitigating these risks. To explore the potential generalizability across languages and disorders, we validated our approach on a dataset of Spanish-speaking Parkinson's disease patients, leveraging pretrained models from healthy English-speaking datasets, and demonstrated that careful pretraining on large-scale task-specific datasets can maintain favorable accuracy under DP constraints. A comprehensive fairness analysis revealed minimal gender bias at reasonable privacy levels but underscored the need for addressing age-related disparities. Our results establish that DP can balance privacy and utility in speech disorder detection, while highlighting unique challenges in privacy-fairness trade-offs for speech data. This provides a foundation for refining DP methodologies and improving fairness across diverse patient groups in real-world deployments.

Abner Hernandez, Paula Andrea Pérez-Toro, Juan Camilo Vásquez-Correa et al.

Collecting speech data is an important step in training speech recognition systems and other speech-based machine learning models. However, the issue of privacy protection is an increasing concern that must be addressed. The current study investigates the use of voice conversion as a method for anonymizing voices. In particular, we train several voice conversion models using self-supervised speech representations including Wav2Vec2.0, Hubert and UniSpeech. Converted voices retain a low word error rate within 1% of the original voice. Equal error rate increases from 1.52% to 46.24% on the LibriSpeech test set and from 3.75% to 45.84% on speakers from the VCTK corpus which signifies degraded performance on speaker verification. Lastly, we conduct experiments on dysarthric speech data to show that speech features relevant to articulation, prosody, phonation and phonology can be extracted from anonymized voices for discriminating between healthy and pathological speech.

Daiqi Liu, Lukas Mulzer, Md Hasan et al.

Segmenting vocal tract articulators in real-time MRI (rtMRI) is a challenging dynamic image segmentation problem characterized by low contrast, rapid motion, and limited spatial resolution. However, while rtMRI acquisitions may provide synchronized acoustic signals, existing methods discard this information, and the few multimodal approaches that incorporate audio cannot be deployed when audio is unavailable. We propose a three-stage framework that leverages acoustic and phonological supervision during training while requiring only the rtMRI image at inference: phonological representations are converted into spatial bounding-box priors for articulator localization, visual and acoustic encoders are aligned via dual-level cross-modal contrastive pretraining, and the learned representations are fused through a cross-attention decoder, effectively transferring multimodal knowledge into a single-modality inference pipeline. Evaluated on 75-Speaker~Annot-16 and USC-TIMIT datasets, our method outperforms existing unimodal and multimodal methods, demonstrating that multimodal supervision provides transferable benefits for precise and clinically deployable vocal tract segmentation.

Andreas Maier, Md Hasan, Paulina Conrad et al.

We describe an approach for estimating the statistical risk of deep networks that contain a mix of learned and known operators. Building on the maximal training error bounds previously established for known operator learning, we derive a deep risk estimator that connects the expected error of a layered network to the size of the training sample. The estimator decomposes the total risk into a sum over learned layers; every known operator contributes zero to this sum, while every learned layer adds an approximation term inspired by Barron's classic work and an estimation term that decreases with the number of training samples. We are able to show that the bound shrinks whenever a learned layer is replaced by a known operator and that the corresponding sample requirement scales with the number of trainable parameters of the layer that is replaced. As an application, we use computed tomography as an example and compare an operator-aware filtered backprojection network with a fully connected substitute that collapses the entire reconstruction pipeline into a single learned dense matrix. The predicted parameter ratio coincides with the structural sparsity that the analytic decomposition into a circulant filter and a sparse backprojection exposes. We confirm the predicted scaling on CPU at small image scale and on GPU at medium image scale, all on the same scaling law. Beyond CT reconstruction, the estimator applies to physics-informed neural networks that hardcode a known physical operation in its architecture, and we expect the result to be of interest for a broad community working on operator-aware deep learning. Calibrating the per-layer constants on each sweep yields a bound that tracks the empirical test MSE within a factor of two at every training-set size, so the estimator can be inverted to predict how many training samples are required to reach a target error.

Tobias Weise, Philipp Klumpp, Kubilay Can Demir et al.

This paper introduces a novel combination of two tasks, previously treated separately: acoustic-to-articulatory speech inversion (AAI) and phoneme-to-articulatory (PTA) motion estimation. We refer to this joint task as acoustic phoneme-to-articulatory speech inversion (APTAI) and explore two different approaches, both working speaker- and text-independently during inference. We use a multi-task learning setup, with the end-to-end goal of taking raw speech as input and estimating the corresponding articulatory movements, phoneme sequence, and phoneme alignment. While both proposed approaches share these same requirements, they differ in their way of achieving phoneme-related predictions: one is based on frame classification, the other on a two-staged training procedure and forced alignment. We reach competitive performance of 0.73 mean correlation for the AAI task and achieve up to approximately 87% frame overlap compared to a state-of-the-art text-dependent phoneme force aligner.

Abner Hernandez, Tomás Arias-Vergara, Daiqi Liu et al.

Phonological features provide a language-general and linguistically grounded representation of speech. We present PhonoQ-2.0, a multilingual frame-level phonological feature recognizer built on self-supervised speech models. The system directly predicts a structured 22-dimensional feature vector per frame encoding manner, vowel quality, place, and voicing, instead of deriving features from phoneme outputs. To ensure phonologically coherent predictions, we introduce a manner-conditioned gating mechanism that activates valid feature groups. Evaluated across multiple languages and corpora, PhonoQ-2.0 achieves an average macro-F1 of 91.3% in-domain and 88.9% out-of-domain. Compared to a strong CTC phoneme baseline, it delivers consistent gains of +8.8 F1 in-domain and +8.6 out-of-domain on average. In unseen-language evaluation, PhonoQ-2.0 improves macro-F1 from 66.9% to 73.6% (+6.7 on average), with gains of up to +10.8 points.

Daiqi Liu, Tomás Arias-Vergara, Jana Hutter et al.

Accurate classification of articulatory-phonological features plays a vital role in understanding human speech production and developing robust speech technologies, particularly in clinical contexts where targeted phonemic analysis and therapy can improve disease diagnosis accuracy and personalized rehabilitation. In this work, we propose a multimodal deep learning framework that combines real-time magnetic resonance imaging (rtMRI) and speech signals to classify three key articulatory dimensions: manner of articulation, place of articulation, and voicing. We perform classification on 15 phonological classes derived from the aforementioned articulatory dimensions and evaluate the system with four audio/vision configurations: unimodal rtMRI, unimodal audio signals, multimodal middle fusion, and contrastive learning-based audio-vision fusion. Experimental results on the USC-TIMIT dataset show that our contrastive learning-based approach achieves state-of-the-art performance, with an average F1-score of 0.81, representing an absolute increase of 0.23 over the unimodal baseline. The results confirm the effectiveness of contrastive representation learning for multimodal articulatory analysis. Our code and processed dataset will be made publicly available at https://github.com/DaE-plz/AC_Contrastive_Phonology to support future research.

Kashaf Gulzar, Korbinian Riedhammer, Elmar Nöth et al.

Speech-based detection of cognitive impairment (CI) offers a promising non-invasive approach for early diagnosis, yet performance disparities across demographic and clinical subgroups remain underexplored, raising concerns around fairness and generalizability. This study presents a systematic bias analysis of acoustic-based CI and depression classification using the DementiaBank Pitt Corpus. We compare traditional acoustic features (MFCCs, eGeMAPS) with contextualized speech embeddings from Wav2Vec 2.0 (W2V2), and evaluate classification performance across gender, age, and depression-status subgroups. For CI detection, higher-layer W2V2 embeddings outperform baseline features (UAR up to 80.6\%), but exhibit performance disparities; specifically, females and younger participants demonstrate lower discriminative power (\(AUC\): 0.769 and 0.746, respectively) and substantial specificity disparities (\(Δ_{spec}\) up to 18\% and 15\%, respectively), leading to a higher risk of misclassifications than their counterparts. These disparities reflect representational biases, defined as systematic differences in model performance across demographic or clinical subgroups. Depression detection within CI subjects yields lower overall performance, with mild improvements from low and mid-level W2V2 layers. Cross-task generalization between CI and depression classification is limited, indicating that each task depends on distinct representations. These findings emphasize the need for fairness-aware model evaluation and subgroup-specific analysis in clinical speech applications, particularly in light of demographic and clinical heterogeneity in real-world applications.

Soroosh Tayebi Arasteh, Tomas Arias-Vergara, Paula Andrea Perez-Toro et al.

Integration of speech into healthcare has intensified privacy concerns due to its potential as a non-invasive biomarker containing individual biometric information. In response, speaker anonymization aims to conceal personally identifiable information while retaining crucial linguistic content. However, the application of anonymization techniques to pathological speech, a critical area where privacy is especially vital, has not been extensively examined. This study investigates anonymization's impact on pathological speech across over 2,700 speakers from multiple German institutions, focusing on privacy, pathological utility, and demographic fairness. We explore both deep-learning-based and signal processing-based anonymization methods. We document substantial privacy improvements across disorders-evidenced by equal error rate increases up to 1933%, with minimal overall impact on utility. Specific disorders such as Dysarthria, Dysphonia, and Cleft Lip and Palate experience minimal utility changes, while Dysglossia shows slight improvements. Our findings underscore that the impact of anonymization varies substantially across different disorders. This necessitates disorder-specific anonymization strategies to optimally balance privacy with diagnostic utility. Additionally, our fairness analysis reveals consistent anonymization effects across most of the demographics. This study demonstrates the effectiveness of anonymization in pathological speech for enhancing privacy, while also highlighting the importance of customized and disorder-specific approaches to account for inversion attacks.

Julian Oelhaf, Georg Kordowich, Changhun Kim et al.

Germany's transition to a renewable energy-based power system is reshaping grid operations, requiring advanced monitoring and control to manage decentralized generation. Machine learning (ML) has emerged as a powerful tool for power system protection, particularly for fault detection (FD) and fault line identification (FLI) in transmission grids. However, ML model reliability depends on data quality and availability. Data sparsity resulting from sensor failures, communication disruptions, or reduced sampling rates poses a challenge to ML-based FD and FLI. Yet, its impact has not been systematically validated prior to this work. In response, we propose a framework to assess the impact of data sparsity on ML-based FD and FLI performance. We simulate realistic data sparsity scenarios, evaluate their impact, derive quantitative insights, and demonstrate the effectiveness of this evaluation strategy by applying it to an existing ML-based framework. Results show the ML model remains robust for FD, maintaining an F1-score of 0.999 $\pm$ 0.000 even after a 50x data reduction. In contrast, FLI is more sensitive, with performance decreasing by 55.61% for missing voltage measurements and 9.73% due to communication failures at critical network points. These findings offer actionable insights for optimizing ML models for real-world grid protection. This enables more efficient FD and supports targeted improvements in FLI.

Chin-Jou Li, Eunjung Yeo, Kwanghee Choi et al. · cmu

Automatic speech recognition (ASR) for dysarthric speech remains challenging due to data scarcity, particularly in non-English languages. To address this, we fine-tune a voice conversion model on English dysarthric speech (UASpeech) to encode both speaker characteristics and prosodic distortions, then apply it to convert healthy non-English speech (FLEURS) into non-English dysarthric-like speech. The generated data is then used to fine-tune a multilingual ASR model, Massively Multilingual Speech (MMS), for improved dysarthric speech recognition. Evaluation on PC-GITA (Spanish), EasyCall (Italian), and SSNCE (Tamil) demonstrates that VC with both speaker and prosody conversion significantly outperforms the off-the-shelf MMS performance and conventional augmentation techniques such as speed and tempo perturbation. Objective and subjective analyses of the generated data further confirm that the generated speech simulates dysarthric characteristics.

Ismael Agchar, Ilja Baumann, Franziska Braun et al.

In recent years, machine learning, and in particular generative adversarial neural networks (GANs) and attention-based neural networks (transformers), have been successfully used to compose and generate music, both melodies and polyphonic pieces. Current research focuses foremost on style replication (eg. generating a Bach-style chorale) or style transfer (eg. classical to jazz) based on large amounts of recorded or transcribed music, which in turn also allows for fairly straight-forward "performance" evaluation. However, most of these models are not suitable for human-machine co-creation through live interaction, neither is clear, how such models and resulting creations would be evaluated. This article presents a thorough review of music representation, feature analysis, heuristic algorithms, statistical and parametric modelling, and human and automatic evaluation measures, along with a discussion of which approaches and models seem most suitable for live interaction.

Julian Oelhaf, Georg Kordowich, Changhun Kim et al.

The increasing integration of distributed energy resources (DERs), particularly renewables, poses significant challenges for power system protection, with fault classification (FC) and fault localization (FL) being among the most critical tasks. Conventional protection schemes, based on fixed thresholds, cannot reliably identify and localize short circuits with the increasing complexity of the grid under dynamic conditions. Machine learning (ML) offers a promising alternative; however, systematic benchmarks across models and settings remain limited. This work presents, for the first time, a comparative benchmarking study of classical ML models for FC and FL in power system protection based on EMT data. Using voltage and current waveforms segmented into sliding windows of 10 ms to 50 ms, we evaluate models under realistic real-time constraints. Performance is assessed in terms of accuracy, robustness to window size, and runtime efficiency. The best-performing FC model achieved an F1 score of 0.992$\pm$0.001, while the top FL model reached an R2 of 0.806$\pm$0.008 with a mean processing time of 0.563 ms.

Abner Hernandez, Tomás Arias Vergara, Andreas Maier et al.

We investigate the use of large language models (LLMs) as post-processing modules for automatic speech recognition (ASR), focusing on their ability to perform error correction for disordered speech. In particular, we propose confidence-informed prompting, where word-level uncertainty estimates are embedded directly into LLM training to improve robustness and generalization across speakers and datasets. This approach directs the model to uncertain ASR regions and reduces overcorrection. We fine-tune a LLaMA 3.1 model and compare our approach to both transcript-only fine-tuning and post hoc confidence-based filtering. Evaluations show that our method achieves a 10% relative WER reduction compared to naive LLM correction on the Speech Accessibility Project spontaneous speech and a 47% reduction on TORGO, demonstrating the effectiveness of confidence-aware fine-tuning for impaired speech.

Lukas Buess, Jan Geier, David Bani-Harouni et al.

Spoken communication plays a central role in clinical workflows. In radiology, for example, most reports are created through dictation. Yet, nearly all medical AI systems rely exclusively on written text. In this work, we address this gap by exploring the feasibility of learning visual-language representations directly from spoken radiology reports. Specifically, we synthesize a large-scale dataset (Speech-RATE) of spoken radiology reports and train SpeechCT-CLIP, a contrastive model that aligns speech and 3D CT volumes in a shared representation space. While naive speech-based models underperform compared to text-trained counterparts, we show that knowledge distillation from a pretrained text-image CLIP model effectively transfers semantic alignment capabilities from text to speech, substantially narrowing this gap. Experiments demonstrate improved zero-shot classification F1 from 0.623 to 0.705, recovering 88% of the performance difference, and strong retrieval results without requiring text at inference. These findings highlight speech as a practical alternative to text in multimodal pretraining and open the door to voice-driven diagnostic support tools in clinical practice.

Daiqi Liu, Tomás Arias-Vergara, Johannes Enk et al.

Accurately segmenting articulatory structures in real-time magnetic resonance imaging (rtMRI) remains challenging, as most existing methods rely almost entirely on visual cues. Yet synchronized acoustic and phonological signals provide complementary context that can enrich visual information and improve precision. In this paper, we introduce VocSegMRI, a multimodal framework that integrates video, audio, and phonological inputs through cross-attention fusion for dynamic feature alignment. To further enhance cross-modal representation, we incorporate a contrastive learning objective that improves segmentation performance even when the audio modality is unavailable at inference. Evaluated on a sub-set of USC-75 rtMRI dataset, our approach achieves state-of-the-art performance, with a Dice score of 0.95 and a 95th percentile Hausdorff Distance (HD_95) of 4.20 mm, outperforming both unimodal and multimodal baselines. Ablation studies confirm the contributions of cross-attention and contrastive learning to segmentation precision and robustness. These results highlight the value of integrative multimodal modeling for accurate vocal tract analysis.

Philipp Klumpp, Tomás Arias-Vergara, Paula Andrea Pérez-Toro et al.

Current state of the art acoustic models can easily comprise more than 100 million parameters. This growing complexity demands larger training datasets to maintain a decent generalization of the final decision function. An ideal dataset is not necessarily large in size, but large with respect to the amount of unique speakers, utilized hardware and varying recording conditions. This enables a machine learning model to explore as much of the domain-specific input space as possible during parameter estimation. This work introduces Common Phone, a gender-balanced, multilingual corpus recorded from more than 11.000 contributors via Mozilla's Common Voice project. It comprises around 116 hours of speech enriched with automatically generated phonetic segmentation. A Wav2Vec 2.0 acoustic model was trained with the Common Phone to perform phonetic symbol recognition and validate the quality of the generated phonetic annotation. The architecture achieved a PER of 18.1 % on the entire test set, computed with all 101 unique phonetic symbols, showing slight differences between the individual languages. We conclude that Common Phone provides sufficient variability and reliable phonetic annotation to help bridging the gap between research and application of acoustic models.

Philipp Klumpp, Tomás Arias-Vergara, Juan Camilo Vásquez-Correa et al.

As one of the most prevalent neurodegenerative disorders, Parkinson's disease (PD) has a significant impact on the fine motor skills of patients. The complex interplay of different articulators during speech production and realization of required muscle tension become increasingly difficult, thus leading to a dysarthric speech. Characteristic patterns such as vowel instability, slurred pronunciation and slow speech can often be observed in the affected individuals and were analyzed in previous studies to determine the presence and progression of PD. In this work, we used a phonetic recognizer trained exclusively on healthy speech data to investigate how PD affected the phonetic footprint of patients. We rediscovered numerous patterns that had been described in previous contributions although our system had never seen any pathological speech previously. Furthermore, we could show that intermediate activations from the neural network could serve as feature vectors encoding information related to the disease state of individuals. We were also able to directly correlate the expert-rated intelligibility of a speaker with the mean confidence of phonetic predictions. Our results support the assumption that pathological data is not necessarily required to train systems that are capable of analyzing PD speech.