# Neue Publikationen in Ziel-Venues – 2026-10-26 Neue Werke: 24 · davon mit DE-Affiliation: 24 Konferenz-Durchgang ab Jahrgang 2025 · letzter Lauf: 2026-10-26T04:49:22 ## IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) — 24 neu · _conference_ ### Binary Modulation on Conjugate-Reciprocal Zeros (MOCZ) with List Decoding for Unknown Channel Length ⚑ DE - **DOI:** [10.1109/icassp55912.2026.11460430](https://doi.org/10.1109/icassp55912.2026.11460430) - **Metadaten:** Erschienen: 2026-04-21 · S. 21361-21365 · OpenAlex seit 2026-04-22 - **DE-Institutionen:** Offenburg University of Applied Sciences - **Autoren:** - Stephan Pfletschinger — Offenburg University of Applied Sciences (DE) - Christian Reich — Offenburg University of Applied Sciences (DE) - Mònica Navarro — Centre Tecnologic de Telecomunicacions de Catalunya (ES) - **Topics:** Advanced Wireless Communication Techniques; Error Correcting Code Techniques; PAPR reduction in OFDM - **Keywords:** list decoding; noncoherent communication; multipath channels; LDPC codes; belief propagation decoding; channel impulse response - **Abstract:** MOCZ (modulation on conjugate-reciprocal zeros) has been introduced as a non-coherent communication scheme for short packets over multi-path channels which does not require channel estimation at the receiver. In this paper, we combine MOCZ with state-of-the-art channel codes and address the problem of the unknown length of the channel impulse response by list decoding. Two criteria for the message selection from the list are evaluated and we observe that LDPC coding with the usual belief-propagation decoding performs near-optimum for realistic packet sizes. ### Fine-Tuning Bigvgan-V2 for Robust Musical Tuning Preservation ⚑ DE - **DOI:** [10.1109/icassp55912.2026.11460697](https://doi.org/10.1109/icassp55912.2026.11460697) - **Metadaten:** Erschienen: 2026-04-21 · S. 15727-15731 · OpenAlex seit 2026-04-22 - **DE-Institutionen:** International Audio Laboratories Erlangen - **Autoren:** - Hans-Ulrich Berendes — International Audio Laboratories Erlangen (DE) - Ben Maman — International Audio Laboratories Erlangen (DE) - Meinard Müller — International Audio Laboratories Erlangen (DE) - **Topics:** Music and Audio Processing; Music Technology and Sound Studies; Speech Recognition and Synthesis - **Keywords:** neural vocoder; pitch shifting; fine-tuning; sound quality; music synthesis - **Abstract:** Recent work has shown that neural vocoders can be prone to musical tuning bias, shifting the pitch reference of input signals, and reducing audio quality in cases of non-standard tuning. In this work, we investigate methods to mitigate this issue by fine-tuning, using the popular BigVGAN-V2 vocoder as a case study. Our goal is to make the vocoder reliable across a wide range of tunings and to examine whether reducing tuning bias also leads to more consistent audio quality. To this end, we explore two training approaches: one based on recordings with natural tuning variation, and another using data created through pitch-shift augmentation to replicate tuning diversity. The results demonstrate that targeted adaptation of BigVGAN-V2 mitigates tuning bias and improves audio quality for signals with non-standard tuning. In particular, our fine-tuned model with 80 mel bands matches the tuning preservation of the higher-resolution 128-band version, providing an efficient and robust solution for music synthesis across diverse tuning conditions. ### Sim-MSTNet: Sim2real Based Multi-Task Spatiotemporal Network Traffic Forecasting ⚑ DE - **DOI:** [10.1109/icassp55912.2026.11463013](https://doi.org/10.1109/icassp55912.2026.11463013) - **Metadaten:** Erschienen: 2026-04-21 · S. 1226-1230 · OpenAlex seit 2026-04-22 - **DE-Institutionen:** Hochschule Bielefeld - **Autoren:** - Hui Ma — Xinjiang University (CN) - Qingzhong Li — Xinjiang University (CN) - Jin Wang — Xinjiang University (CN) - Jie Wu — Xinjiang University (CN) - Shaoyu Dou — Tongji University (CN) - Li Feng — Hochschule Bielefeld (DE) - Xinjun Pei — Xinjiang University (CN) - **Topics:** Traffic Prediction and Management Techniques; Software-Defined Networks and 5G; Internet Traffic Analysis and Secure E-voting - **Keywords:** network traffic prediction; multi-task learning; sim-to-real transfer; synthetic data generation; domain randomization; negative transfer; attention mechanism; dynamic loss weighting - **Abstract:** Network traffic forecasting plays a crucial role in intelligent network operations, but existing techniques often perform poorly when faced with limited data. Additionally, multi-task learning methods struggle with task imbalance and negative transfer, especially when modeling various service types. To overcome these challenges, we propose Sim-MSTNet, a multi-task spatiotemporal network traffic forecasting model based on the sim2real approach. Our method leverages a simulator to generate synthetic data, effectively addressing the issue of poor generalization caused by data scarcity. By employing a domain randomization technique, we reduce the distributional gap between synthetic and real data through bi-level optimization of both sample weighting and model training. Moreover, Sim-MSTNet incorporates attention-based mechanisms to selectively share knowledge between tasks and applies dynamic loss weighting to balance task objectives. Extensive experiments on two open-source datasets show that Sim-MSTNet consistently outperforms state-of-the-art baselines, achieving enhanced accuracy and generalization. ### Reference Microphone Selection for Guided Source Separation Based on The Normalized L-P Norm ⚑ DE - **DOI:** [10.1109/icassp55912.2026.11463742](https://doi.org/10.1109/icassp55912.2026.11463742) - **Metadaten:** Erschienen: 2026-04-21 · S. 14712-14716 · OpenAlex seit 2025-11-05 - **DE-Institutionen:** Carl von Ossietzky Universität Oldenburg - **Autoren:** - Anselm Lohmann — Carl von Ossietzky Universität Oldenburg (DE) - Tomohiro Nakatani — NTT (Japan) (JP) - Rintaro Ikeshita — NTT (Japan) (JP) - Marc Delcroix — NTT (Japan) (JP) - Shoko Araki — NTT (Japan) (JP) - Simon Doclo — Carl von Ossietzky Universität Oldenburg (DE) - **Topics:** Speech and Audio Processing; Speech Recognition and Synthesis; Music and Audio Processing - **Keywords:** distant speech recognition; speech enhancement; signal-to-noise ratio - **Abstract:** Guided Source Separation (GSS) is a popular front-end for distant automatic speech recognition (ASR) systems using spatially distributed microphones. When considering spatially distributed microphones, the choice of reference microphone may have a large influence on the quality of the output signal and the downstream ASR performance. In GSS-based speech enhancement, reference microphone selection is typically performed using the signal-to-noise ratio (SNR), which is optimal for noise reduction but may neglect differences in early-to-late reverberation ratio (ELR) across microphones. In this paper, we propose two reference microphone selection methods for GSS-based speech enhancement that are based on the normalized ℓp-norm, either using only the normalized ℓp-norm or combining the normalized ℓp-norm and the SNR to account for both differences in SNR and ELR across microphones. Experimental evaluation using a CHiME-8 distant ASR system shows that the proposed ℓp-norm-based methods outperform the baseline method, reducing the macro-average word error rate. ### Demystifying the Roles of LLM Layers in Retrieval, Knowledge, and Reasoning ⚑ DE - **DOI:** [10.1109/icassp55912.2026.11462157](https://doi.org/10.1109/icassp55912.2026.11462157) - **Metadaten:** Erschienen: 2026-04-21 · S. 16792-16796 · OpenAlex seit 2025-10-10 - **DE-Institutionen:** University of Tübingen; Max Planck Institute for Intelligent Systems - **Autoren:** - Xinyuan Song — Emory University (US) - Keyu Wang — University of Tübingen (DE) - PengXiang Li — Hong Kong Polytechnic University (HK) - Lu Yin — University of Surrey (GB) - Shiwei Liu — Max Planck Institute for Intelligent Systems (DE) - **Topics:** Topic Modeling; Advanced Neural Network Applications; Multimodal Machine Learning Applications - **Keywords:** large language models; layer pruning; knowledge distillation; retrieval; reasoning; long-range coherence; model compression; representation learning - **Abstract:** Recent studies suggest that the deeper layers of Large Language Models (LLMs) contribute little to representation learning and can often be removed without significant performance loss. However, such claims are typically drawn from narrow evaluations and may overlook important aspects of model behavior. In this work, we present a systematic study of depth utilization across diverse dimensions, including evaluation protocols, task categories, and model architectures. Our analysis confirms that very deep layers are generally less effective than earlier ones, but their contributions vary substantially with the evaluation setting. Under likelihood-based metrics without generation, pruning most layers preserves performance, with only the initial few being critical. By contrast, generation-based evaluation uncovers indispensable roles for middle and deeper layers in enabling reasoning and maintaining long-range coherence. We further find that knowledge and retrieval are concentrated in shallow components, whereas reasoning accuracy relies heavily on deeper layers—yet can be reshaped through distillation. These results highlight that depth usage in LLMs is highly heterogeneous and context-dependent, underscoring the need for task-, metric-, and model-aware perspectives in both interpreting and compressing large models. ### Speech Quality-Based Localization of Low-Quality Speech and Text-to-Speech Synthesis Artefacts ⚑ DE - **DOI:** [10.1109/icassp55912.2026.11464613](https://doi.org/10.1109/icassp55912.2026.11464613) - **Metadaten:** Erschienen: 2026-04-21 · S. 15002-15006 · OpenAlex seit 2026-04-22 - **DE-Institutionen:** Paderborn University - **Autoren:** - Michael Kuhlmann — Paderborn University (DE) - Alexander Werning — Paderborn University (DE) - Thilo von Neumann — Paderborn University (DE) - Reinhold Haeb‐Umbach — Paderborn University (DE) - **Topics:** Speech Recognition and Synthesis; Speech and Audio Processing; Image and Video Quality Assessment - **Keywords:** speech quality assessment; model regularization; anti-spoofing - **Abstract:** A large number of works view the automatic assessment of speech from an utterance- or system-level perspective. While such approaches are good in judging overall quality, they cannot adequately explain why a certain score was assigned to an utterance. frame-level scores can provide better interpretability, but models predicting them are harder to tune and regularize since no strong targets are available during training. In this work, we show that utterance-level speech quality predictors can be regularized with a segment-based consistency constraint which notably reduces frame-level stochasticity. We then demonstrate two applications involving frame-level scores: The partial spoof scenario and the detection of synthesis artefacts in two state-of-the-art text-to-speech systems. For the latter, we perform listening tests and confirm that listeners rate segments to be of poor quality more often in the set defined by low frame-level scores than in a random control set. ### X-Ray Near-Field Holotomography Reconstruction Using Implicit Neural Representations ⚑ DE - **DOI:** [10.1109/icassp55912.2026.11463026](https://doi.org/10.1109/icassp55912.2026.11463026) - **Metadaten:** Erschienen: 2026-04-21 · S. 21927-21931 · OpenAlex seit 2026-04-22 - **DE-Institutionen:** Deutsches Elektronen-Synchrotron DESY; Helmholtz-Zentrum Hereon - **Autoren:** - Johannes Gruen — Deutsches Elektronen-Synchrotron DESY (DE) - Sebastian Eberle — Deutsches Elektronen-Synchrotron DESY (DE) - Imke Greving — Helmholtz-Zentrum Hereon (DE) - Silja Flenner — Helmholtz-Zentrum Hereon (DE) - Martin Burger — Deutsches Elektronen-Synchrotron DESY (DE) - Christian G. Schroer — Deutsches Elektronen-Synchrotron DESY (DE) - Johannes Hagemann — Deutsches Elektronen-Synchrotron DESY (DE) - **Topics:** Advanced X-ray Imaging Techniques; Digital Holography and Microscopy; X-ray Diffraction in Crystallography - **Keywords:** implicit neural representations; phase retrieval; tomographic reconstruction; 3D consistency; data compression; synchrotron X-ray imaging - **Abstract:** X-ray near-field holotomography provides non-destructive, in situ 3D visualization of specimen interiors at nanometer-scale resolution. Reconstruction traditionally involves two separate steps: first retrieving the projected phase for different rotation angles, then applying tomographic reconstruction to obtain a 3D volume from 2D projections. Both steps are ill-posed inverse problems and separating them leads to information loss, due to reconstruction errors. Recent advances in implicit neural representations (INRs) have demonstrated remarkable capabilities in scene rendering and tomographic reconstruction. In this work, we propose a unified INR-based framework that jointly solves the phase retrieval and tomographic reconstruction problems. This joint formulation enforces 3D consistency, resulting in significantly improved phase, absorption, and volumetric reconstructions. Moreover, INRs provide substantial data compression. This compression reduces storage requirements by 95%, which is particularly important with the advent of fourth-generation synchrotron sources and the corresponding growth in data volume. ### Matching Reverberant Speech Through Learned Acoustic Embeddings ⚑ DE - **DOI:** [10.1109/icassp55912.2026.11461386](https://doi.org/10.1109/icassp55912.2026.11461386) - **Metadaten:** Erschienen: 2026-04-21 · S. 14577-14581 · OpenAlex seit 2026-04-22 - **DE-Institutionen:** Friedrich-Alexander-Universität Erlangen-Nürnberg; International Audio Laboratories Erlangen; Fraunhofer Institute for Integrated Circuits - **Autoren:** - Philipp Götz — Friedrich-Alexander-Universität Erlangen-Nürnberg (DE); International Audio Laboratories Erlangen (DE); Fraunhofer Institute for Integrated Circuits (DE) - Gloria Dal Santo — Aalto University (FI) - Sebastian J. Schlecht — Friedrich-Alexander-Universität Erlangen-Nürnberg (DE) - Vesa Välimäki — Aalto University (FI) - Emanuël A. P. Habets — Friedrich-Alexander-Universität Erlangen-Nürnberg (DE); International Audio Laboratories Erlangen (DE); Fraunhofer Institute for Integrated Circuits (DE) - **Topics:** Speech and Audio Processing; Speech Recognition and Synthesis; Music and Audio Processing - **Keywords:** reverberation; reverberant speech; feedback delay networks; acoustic embeddings; room acoustics - **Abstract:** Reverberation conveys critical acoustic cues about the environment, supporting spatial awareness and immersion. For auditory augmented reality (AAR) systems, generating perceptually plausible real-time reverberation remains a key challenge, especially when explicit acoustic measurements are unavailable. We address this by formulating blind estimation of artificial reverberation parameters as a reverberant signal matching task, leveraging a learned room-acoustic prior. Furthermore, we propose a feedback delay network (FDN) structure that reproduces both frequency-dependent decay times and the direct-to-reverberation ratio of a target space. Experimental evaluation against a leading automatic FDN tuning method shows improvements in estimated room-acoustic parameters and in the perceptual plausibility of artificial reverberant speech. These results highlight the potential of our approach for efficient, perceptually consistent reverberation rendering in AAR applications. ### Physics-Informed Video Diffusion for Shallow Water Equations ⚑ DE - **DOI:** [10.1109/icassp55912.2026.11462239](https://doi.org/10.1109/icassp55912.2026.11462239) - **Metadaten:** Erschienen: 2026-04-21 · S. 13242-13246 · OpenAlex seit 2026-04-22 - **DE-Institutionen:** Ludwig-Maximilians-Universität München; Huawei German Research Center - **Autoren:** - Yang Bai — Ludwig-Maximilians-Universität München (DE) - George Eskandar — Huawei German Research Center (DE) - Ziyuan Liu — Huawei German Research Center (DE) - Gitta Kutyniok — Ludwig-Maximilians-Universität München (DE) - **Topics:** Generative Adversarial Networks and Image Synthesis; Computer Graphics and Visualization Techniques; Model Reduction and Neural Networks - **Keywords:** shallow water equations; fluid dynamics simulation; generative video models; topography; physical plausibility - **Abstract:** Traditional fluid dynamics simulation pipelines combine numerical solvers with rendering, producing highly realistic results but at considerable computational cost. Diffusion-based generative video models offer a faster alternative, yet often ignore physical laws and thus fail to capture consistent dynamics. We propose a physics-informed video diffusion framework that jointly generates visual outputs and physical states. Unlike prior two-stage approaches that first simulate the physical variables and then render, we directly integrate physics constraints into the generative process, enabling simultaneous prediction of physical states and realistic videos without a separate rendering step. Built on the two-dimensional shallow water equations with terrain topography, our method produces temporally coherent water flow while maintaining physical plausibility. Experiments show that it outperforms purely data-driven video diffusion baselines in both realism and physical fidelity, while generating videos significantly faster than traditional simulation-plus-rendering pipelines. ### Spatially-Coupled OTFS Systems via Block Markov Superposition Transmission ⚑ DE - **DOI:** [10.1109/icassp55912.2026.11464769](https://doi.org/10.1109/icassp55912.2026.11464769) - **Metadaten:** Erschienen: 2026-04-21 · S. 21366-21370 · OpenAlex seit 2026-04-22 - **DE-Institutionen:** Technische Universität Berlin - **Autoren:** - Jiayi Yang — Sun Yat-sen University (CN) - Shuangyang Li — Technische Universität Berlin (DE) - Qianfan Wang — City University of Hong Kong (HK) - Linqi Song — City University of Hong Kong (HK) - Bo Ai — Beijing Jiaotong University (CN) - Xiao Ma — Sun Yat-sen University (CN) - Giuseppe Caire — Technische Universität Berlin (DE) - **Topics:** PAPR reduction in OFDM; Advanced Wireless Communication Techniques; Error Correcting Code Techniques - **Keywords:** orthogonal time frequency space; spatially coupled codes; LDPC codes; decoding threshold; extrinsic information transfer; EXIT charts; doubly selective channels; high-mobility channels - **Abstract:** This paper investigates a novel spatially coupled orthogonal time frequency space (SC-OTFS) transmission scheme by integrating block Markov superposition transmission (BMST) with OTFS modulation. We first statistically analyze the distribution of detector mutual information across coded blocks and establish a decoding-threshold-based lower bound to characterize the asymptotic performance of conventional LDPC-coded OTFS systems. The analysis reveals that increasing the code length yields only marginal coding gains in typical doubly-selective channels with limited resolvable paths. Motivated by this observation, we propose an SC-OTFS system in which a portion of each previously coded block is superimposed onto the current block and subsequently modulated using OTFS, thereby exploiting both coding gain and inter-block diversity gain in high-mobility channels. Furthermore, an extrinsic information transfer (EXIT) chart analysis is developed for the SC-OTFS system, which enables optimization of the critical superposition fraction based on decoding thresholds. Simulation results demonstrate that this optimization leads to a significant performance improvement, achieving more than 3 dB gain over conventional 5G LDPC-coded OTFS systems. ### Combined object-based audio and MASA format for enhanced spatial mobile communication ⚑ DE - **DOI:** [10.1109/icassp49660.2025.10889750](https://doi.org/10.1109/icassp49660.2025.10889750) - **Metadaten:** Erschienen: 2025-03-12 · S. 1-5 · OpenAlex seit 2025-10-10 - **DE-Institutionen:** Nokia (Germany) - **Autoren:** - Mikko‐Ville Laitinen — Nokia (Finland) (FI) - Adriana Vasilache — Nokia (Finland) (FI) - Anssi Rämö — Nokia (Finland) (FI) - Jouni Paulus — Nokia (Germany) (DE) - Václav Eksler — VoiceAge (Canada) (CA) - **Topics:** Speech and Audio Processing; Music and Audio Processing; Digital Filter Design and Implementation - **Keywords:** object-based audio; spatial audio; 5G; teleconferencing; mobile communication; computational complexity - **Abstract:** The new 3GPP IVAS codec for immersive voice and audio communication services in 5G networks supports a variety of spatial audio formats beyond mono and stereo voice and audio, and even combinations of them. This paper presents the OMASA combined audio format which provides a joint coding of object-based audio and metadata-assisted spatial audio (MASA). It enables immersive communication experiences, e.g., in teleconferencing scenarios where a mobile end-device captures the environmental audio in MASA format while individual talkers’ voices are captured with their own microphones as separate audio objects. This paper describes coding methods newly developed and optimized depending on the available bitrate and the number of audio objects to efficiently deliver a 3D audio experience at a wide OMASA bitrate range from 13.2 kbps to 512 kbps. The benefits of the presented joint coding of audio formats over separate coding at the same bitrates are presented in terms of subjective evaluation and computational resources. ### PIER: A Novel Metric for Evaluating What Matters in Code-Switching ⚑ DE - **DOI:** [10.1109/icassp49660.2025.10889660](https://doi.org/10.1109/icassp49660.2025.10889660) - **Metadaten:** Erschienen: 2025-03-12 · S. 1-5 · OpenAlex seit 2025-10-10 - **DE-Institutionen:** Karlsruhe Institute of Technology - **Autoren:** - Enes Yavuz Ugan — Karlsruhe Institute of Technology (DE) - Ngoc-Quan Pham — Karlsruhe Institute of Technology (DE) - Leonard Bärmann — Karlsruhe Institute of Technology (DE) - Alex Waibel — Carnegie Mellon University (US) - **Topics:** Speech Recognition and Synthesis; Speech and Audio Processing; Speech and dialogue systems - **Keywords:** code-switching; word error rate; language alternation; embedded languages; connectionist temporal classification; encoder-decoder architecture - **Abstract:** Code-switching, the alternation of languages within a single discourse, presents a significant challenge for Automatic Speech Recognition. Despite the unique nature of the task, performance is commonly measured with established metrics such as Word-Error-Rate (WER). However, in this paper, we question whether these general metrics accurately assess performance on code-switching. Specifically, using both Connectionist-Temporal-Classification and Encoder-Decoder models, we show fine-tuning on non-code-switched data from both matrix and embedded language improves classical metrics on code-switching test sets, although actual code-switched words worsen (as expected). Therefore, we propose Point-of-Interest Error Rate (PIER), a variant of WER that focuses only on specific words of interest. We instantiate PIER on code-switched utterances and show that this more accurately describes the code-switching performance, showing huge room for improvement in future work. This focused evaluation allows for a more precise assessment of model performance, particularly in challenging aspects such as inter-word and intra-word code-switching. ### COCOLA: Coherence-Oriented Contrastive Learning of Musical Audio Representations ⚑ DE - **DOI:** [10.1109/icassp49660.2025.10888146](https://doi.org/10.1109/icassp49660.2025.10888146) - **Metadaten:** Erschienen: 2025-03-12 · S. 1-5 · OpenAlex seit 2025-10-10 - **DE-Institutionen:** University of Stuttgart - **Autoren:** - Ruben Ciranni — Sapienza University of Rome (IT) - Giorgio Mariani — Sapienza University of Rome (IT) - Michele Mancusi — University of Stuttgart (DE) - Emilian Postolache — Ca' Foscari University of Venice (IT) - Giorgio Fabbro — University of Stuttgart (DE) - Emanuele Rodolà — Sapienza University of Rome (IT) - Luca Cosmo — Ca' Foscari University of Venice (IT) - **Topics:** Music and Audio Processing; Music Technology and Sound Studies; Speech and Audio Processing - **Keywords:** contrastive learning; harmonic coherence; generative model evaluation; audio representation learning - **Abstract:** We present COCOLA (Coherence-Oriented Contrastive Learning for Audio), a contrastive learning method for musical audio representations that captures the harmonic and rhythmic coherence between samples. Our method operates at the level of the individual stems composing music tracks and can input features obtained via Harmonic-Percussive Separation (HPS). COCOLA allows an objective evaluation of generative models for music accompaniment generation, which are difficult to benchmark with established metrics. In this regard, we evaluate recent music accompaniment generation models, demonstrating the effectiveness of our proposed method. We release the model checkpoints trained on public datasets containing separate stems (MUSDB18-HQ, MoisesDB, Slakh2100, and CocoChorales). ### Classifying Music-Induced Emotion Using Multi-Modal Ensembles of EEG and Audio Feature Models ⚑ DE - **DOI:** [10.1109/icassp49660.2025.10888818](https://doi.org/10.1109/icassp49660.2025.10888818) - **Metadaten:** Erschienen: 2025-03-12 · S. 1-5 · OpenAlex seit 2025-10-10 - **DE-Institutionen:** Technical University of Munich - **Autoren:** - Philipp Paukner — Technical University of Munich (DE) - Marisa Ripoll — Technical University of Munich (DE) - Dilvan Sabir — Technical University of Munich (DE) - Deniz Onat Erdoğan — Technical University of Munich (DE) - Luca Sacchetto — Technical University of Munich (DE) - Klaus J. Diepold — Technical University of Munich (DE) - **Topics:** Emotion and Mood Recognition; Music and Audio Processing; EEG and Brain-Computer Interfaces - **Keywords:** emotional response to music; EEG; audio features; multimodal ensemble; Audio Spectrogram Transformer; emotion recognition - **Abstract:** In this paper, we present our submission to the EEG-Music Emotion Recognition Challenge at ICASSP 2025. Our work focused on Task 2, where the objective was to classify the emotional state of subjects in a discrete valence-arousal space while they listen to music. Our proposed solution adopts an ensemble approach, integrating electroencephalography (EEG) signals, raw audio signals, and song features. We incorporated a diverse set of models, including Audio Spectrogram Transformer (AST) [1] and a dedicated EEG model. By combining insights from these modalities, we aimed to capture the interplay between music and emotional responses. We achieved a balanced accuracy of 41.34% on held-out data, improving the baseline by more than 11% and thus 2ndplace in the competition. ### Precisely Controllable Neural Speech Synthesis ⚑ DE - **DOI:** [10.1109/icassp49660.2025.10890772](https://doi.org/10.1109/icassp49660.2025.10890772) - **Metadaten:** Erschienen: 2025-03-12 · S. 1-5 · OpenAlex seit 2025-10-10 - **DE-Institutionen:** Technische Universität Dresden - **Autoren:** - Paul Konstantin Krug — Altavo GmbH,Dresden,Germany - Christoph Wagner — Altavo GmbH,Dresden,Germany - Peter Birkholz — Technische Universität Dresden (DE) - Timo Stich — Altavo GmbH,Dresden,Germany - **Topics:** Speech Recognition and Synthesis; Speech and Audio Processing; Phonetics and Phonology Research - **Keywords:** neural speech synthesis; articulatory speech synthesis; controllable speech synthesis; disentangled latent representations; vocal tract modeling; speech intelligibility - **Abstract:** Recent advances in deep learning have significantly improved the quality of speech synthesis, yet these models often suffer from limited controllability and lack of interpretability due to their black-box nature. In contrast, articulatory speech synthesis offers fine-grained control and transparency by simulating sound production based on vocal tract geometry, though it struggles with naturalness and synthesis quality. To bridge these gaps, we propose a novel white-box approach that leverages synthetic articulatory trajectories for neural synthesis, ensuring a fully disentangled, interpretable, and controllable yet high-quality speech synthesis process. Utilizing the VocalTractLab articulatory synthesizer, our method allows the quality of its speech representation to be verified through physical simulation. The proposed system achieves state-of-the-art results in both articulatory and deep articulatory synthesis. To the best of our knowledge, this is the first work to synthesize highly intelligible speech from a purely synthetic articulatory latent representation. ### Latent Diffusion Bridges for Unsupervised Musical Audio Timbre Transfer ⚑ DE - **DOI:** [10.1109/icassp49660.2025.10890708](https://doi.org/10.1109/icassp49660.2025.10890708) - **Metadaten:** Erschienen: 2025-03-12 · S. 1-5 · OpenAlex seit 2025-10-10 - **DE-Institutionen:** University of Stuttgart - **Autoren:** - Michele Mancusi — University of Stuttgart (DE) - Yurii Halychanskyi — University of Illinois Urbana-Champaign (US) - Kin Wai Cheuk — Sony (Japan) (JP) - Eloi Moliner — Aalto University (FI) - Chieh-Hsin Lai — Sony (Japan) (JP) - Stefan Uhlich — University of Stuttgart (DE) - Junghyun Koo — Sony (Japan) (JP) - Marco A. Martínez-Ramírez — Sony (Japan) (JP) - Wei‐Hsiang Liao — Sony (Japan) (JP) - Giorgio Fabbro — University of Stuttgart (DE) - Yuki Mitsufuji — Sony (Japan) (JP) - **Topics:** Music and Audio Processing; Music Technology and Sound Studies; Generative Adversarial Networks and Image Synthesis - **Keywords:** timbre transfer; diffusion models; Gaussian prior; Fréchet Audio Distance - **Abstract:** Music timbre transfer is a challenging task that involves modifying the timbral characteristics of an audio signal while preserving its melodic structure. In this paper, we propose a novel method based on dual diffusion bridges, trained using the CocoChorales Dataset, which consists of unpaired monophonic single-instrument audio data. Each diffusion model is trained on a specific instrument with a Gaussian prior. During inference, a model is designated as the source model to map the input audio to its corresponding Gaussian prior, and another model is designated as the target model to reconstruct the target audio from this Gaussian prior, thereby facilitating timbre transfer. We compare our approach against existing unsupervised timbre transfer models such as VAEGAN and Gaussian Flow Bridges (GFB). Experimental results demonstrate that our method achieves both better Fréchet Audio Distance (FAD) and melody preservation, as reflected by lower pitch distances (DPD) compared to VAEGAN and GFB. Additionally, we discover that the noise level from the Gaussian prior, σ, can be adjusted to control the degree of melody preservation and amount of timbre transferred. ### Digital Operating Mode Classification of Real-World Amateur Radio Transmissions ⚑ DE - **DOI:** [10.1109/icassp49660.2025.10889837](https://doi.org/10.1109/icassp49660.2025.10889837) - **Metadaten:** Erschienen: 2025-03-12 · S. 1-5 · OpenAlex seit 2025-10-10 - **DE-Institutionen:** Georg Simon Ohm University of Applied Sciences Nuremberg - **Autoren:** - Maximilian Bundscherer — Georg Simon Ohm University of Applied Sciences Nuremberg (DE) - Thomas Schmitt — Georg Simon Ohm University of Applied Sciences Nuremberg (DE) - Ilja Baumann — Georg Simon Ohm University of Applied Sciences Nuremberg (DE) - Tobias Bocklet — Georg Simon Ohm University of Applied Sciences Nuremberg (DE) - **Topics:** Wireless Signal Modulation Classification; Cognitive Radio Networks and Spectrum Sensing; Music and Audio Processing - **Keywords:** amateur radio; software-defined radio; spectrogram classification; lightweight neural networks; EfficientNetB0; data augmentation; channel impairments; UHF band - **Abstract:** 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. ### MSA-ASR: Efficient Multilingual Speaker Attribution with frozen ASR Models ⚑ DE - **DOI:** [10.1109/icassp49660.2025.10889116](https://doi.org/10.1109/icassp49660.2025.10889116) - **Metadaten:** Erschienen: 2025-03-12 · S. 1-5 · OpenAlex seit 2025-10-10 - **DE-Institutionen:** Karlsruhe Institute of Technology - **Autoren:** - Thai Binh Nguyen — Karlsruhe Institute of Technology (DE) - Alexander H. Waibel — Carnegie Mellon University (US) - **Topics:** Speech Recognition and Synthesis; Speech and Audio Processing; Authorship Attribution and Profiling - **Keywords:** speaker attribution; speech recognition; multilingual speech recognition ### FARE: A Deep Learning-Based Framework for Radar-Based Face Recognition and Out-of-Distribution Detection ⚑ DE - **DOI:** [10.1109/icassp49660.2025.10889758](https://doi.org/10.1109/icassp49660.2025.10889758) - **Metadaten:** Erschienen: 2025-03-12 · S. 1-5 · OpenAlex seit 2025-10-10 - **DE-Institutionen:** Technical University of Munich - **Autoren:** - Sabri Mustafa Kahya — Technical University of Munich (DE) - Boran Hamdi Sivrikaya — Technical University of Munich (DE) - Muhammet Sami Yavuz — Technical University of Munich (DE) - Eckehard G. Steinbach — Technical University of Munich (DE) - **Topics:** Face recognition and analysis; Advanced SAR Imaging Techniques; Biometric Identification and Security - **Keywords:** out-of-distribution detection; FMCW radar; range-Doppler image; deep learning; triplet loss; autoencoder; 60 GHz radar; short-range radar - **Abstract:** In this work, we propose a novel pipeline for face recognition and out-of-distribution (OOD) detection using shortrange FMCW radar. The proposed system utilizes RangeDoppler and micro Range-Doppler Images. The architecture features a primary path (PP) responsible for the classification of in-distribution (ID) faces, complemented by intermediate paths (IPs) dedicated to OOD detection. The network is trained in two stages: first, the PP is trained using triplet loss to optimize ID face classification. In the second stage, the PP is frozen, and the IPs—comprising simple linear autoen-coder networks—are trained specifically for OOD detection. Using our dataset generated with a 60 GHz FMCW radar, our method achieves an ID classification accuracy of 99.30% and an OOD detection AUROC of 96.91%. ### A Systematic Evaluation of Machine Learning Methods for Fault Detection and Line Identification in Electrical Power Grids ⚑ DE - **DOI:** [10.1109/icassp49660.2025.10890544](https://doi.org/10.1109/icassp49660.2025.10890544) - **Metadaten:** Erschienen: 2025-03-12 · S. 1-5 · OpenAlex seit 2025-10-10 - **DE-Institutionen:** Friedrich-Alexander-Universität Erlangen-Nürnberg - **Autoren:** - Julian Oelhaf — Friedrich-Alexander-Universität Erlangen-Nürnberg (DE) - Georg Kordowich — Friedrich-Alexander-Universität Erlangen-Nürnberg (DE) - Paula Andrea Pérez-Toro — Friedrich-Alexander-Universität Erlangen-Nürnberg (DE) - Tomás Arias‐Vergara — Friedrich-Alexander-Universität Erlangen-Nürnberg (DE) - Andreas Maier — Friedrich-Alexander-Universität Erlangen-Nürnberg (DE) - Johann Jaeger — Friedrich-Alexander-Universität Erlangen-Nürnberg (DE) - Siming Bayer — Friedrich-Alexander-Universität Erlangen-Nürnberg (DE) - **Topics:** Power Systems Fault Detection; Electrical Fault Detection and Protection; Power System Optimization and Stability - **Keywords:** fault detection; electrical power grid; machine learning; short-circuit faults; relay protection; renewable energy integration; real-time fault detection; model comparison - **Abstract:** The integration of renewable energy sources into the electrical grid introduces complex challenges in fault detection and coordination of grid recovery mechanisms. Traditional relay protection systems, which operate based on static rules and predefined thresholds, are inadequate for addressing these challenges, particularly in detecting and isolating faults such as short circuits. Consequently, the conventional methodologies applied to electrical network protection frequently fail to achieve optimal performance in fault detection, especially in terms of adherence to safety standards and the selective limitation of damage. Recent research indicates that machine learning (ML)-based approaches can effectively tackle these issues; however, variations in grid configurations and analysis windows have impeded consistent comparative assessments. In this study, we assess the efficacy of various ML models in detecting electrical faults and pinpointing defective transmission lines within a 10 ms measurement interval—a critical time-frame for real-time operational viability, for the first time. The most effective model attained an F1 score of 0.991±0.018 and demonstrated a processing time of 0.342ms±0.509ms. ### PeT-KeyStAtion: Parameter-efficient Transformer with Keypoint-guided Spatial-temporal Aggregation for Video-based Person Re-identification ⚑ DE - **DOI:** [10.1109/icassp49660.2025.10888219](https://doi.org/10.1109/icassp49660.2025.10888219) - **Metadaten:** Erschienen: 2025-03-12 · S. 1-5 · OpenAlex seit 2025-10-10 - **DE-Institutionen:** University of Bonn - **Autoren:** - Xingan Ma — University of Bonn (DE) - Jinhui Yi — University of Bonn (DE) - Jüergen Gall — University of Bonn (DE) - **Topics:** Video Surveillance and Tracking Methods; Advanced Neural Network Applications; Advanced Image and Video Retrieval Techniques - **Keywords:** video-based person re-identification; lightweight adapters; visual surveillance - **Abstract:** Video-based Person Re-identification (ReID) is crucial in visual surveillance, focusing on matching video snippets of individuals across multiple non-overlapping cameras. Existing methods either conduct ReID at the image level without leveraging temporal information, or employ complex temporal information aggregation techniques, which results in substantial network size and reduced performance efficiency. Recent advances in Vision Transformer (ViT) architectures leverage diverse large-scale datasets alongside sophisticated architectures to achieve enhanced fine-grained feature discrimination. To fully explore the potential of ViT architectures without adding substantial additional modules for video-based ReID, we propose PeT-KeyStAtion: a Parameter-efficient Transformer with Keypoint- guided Spatial-temporal Aggregation using a Spatial-Temporal and Keypoint (STK) Module with lightweight adapters. Our framework effectively captures and aggregates spatial, temporal, and keypoint information with only 11% of the parameters compared to full fine-tuning. Extensive experiments show that our method outperforms state-of-the-art baselines on MARS and iLIDS-VID, and achieves promising performance on LS-VID. ### The Sound of Language: A Bilingual Analysis of Voice Conversion and Text-to-Speech Synthesis ⚑ DE - **DOI:** [10.1109/icassp49660.2025.10888796](https://doi.org/10.1109/icassp49660.2025.10888796) - **Metadaten:** Erschienen: 2025-03-12 · S. 1-5 · OpenAlex seit 2025-10-10 - **DE-Institutionen:** Fraunhofer Institute for Secure Information Technology - **Autoren:** - Jeongeun Choi — Fraunhofer Institute for Secure Information Technology (DE) - Karla Schäfer — Fraunhofer Institute for Secure Information Technology (DE) - Martin Steinebach — Fraunhofer Institute for Secure Information Technology (DE) - **Topics:** Speech Recognition and Synthesis; Speech and Audio Processing; Voice and Speech Disorders - **Keywords:** voice conversion; text-to-speech; deepfake audio; audio quality assessment - **Abstract:** With the rise of audio deepfakes, there is an increasing need for comprehensive studies on their generation methods, especially regarding their quality. Areas such as languages beyond English and Chinese, as well as comparisons between voice conversion (VC) and text-to-speech synthesis (TTS), remain underexplored. In our study, we generated samples in English and German using 10 recent VC and TTS methods, including two publicly accessible online tools. We compared these samples using various evaluation methods to gain insights into their quality across different factors. Our analysis indicates that TTS performs slightly better than VC, with minor differences between English and German data. Interestingly, in VC, the gender of the source speaker has minimal influence on the generated samples. Instead, the cross-gender factor appears to affect VC. For both VC and TTS, the target speaker samples used for generation seem to influence the quality of the generated samples. ### Evaluating Snippet Significance: A Framework for Audio and Text-Based Dialogue Summarization ⚑ DE - **DOI:** [10.1109/icassp49660.2025.10888790](https://doi.org/10.1109/icassp49660.2025.10888790) - **Metadaten:** Erschienen: 2025-03-12 · S. 1-5 · OpenAlex seit 2025-10-10 - **DE-Institutionen:** Technische Hochschule Ingolstadt - **Autoren:** - Anderson Luiz — Technische Hochschule Ingolstadt (DE) - Raviteja Boddu — Technische Hochschule Ingolstadt (DE) - Munir Georges — Technische Hochschule Ingolstadt (DE) - **Topics:** Speech and dialogue systems; Speech Recognition and Synthesis; Topic Modeling - **Keywords:** dialogue summarization; multimodal summarization; audio embeddings; wav2vec 2.0; BERT embeddings; AMI meeting corpus; ICSI meeting corpus; spoken dialogue - **Abstract:** The extraction and summarization of key information from dialogues are essential for improving communication tools and outcomes across various domains. Traditional text-based approaches often fail to capture critical auditory cues such as intonation, pauses, and emphasis, which are vital for understanding the full context of spoken interactions. This paper introduces a novel framework that independently analyzes text and audio modalities, assigning significance scores to dialogue snippets. Evaluations on the AMI [1] and ICSI [2] meeting corpora utilize Wav2Vec 2.0 [3] for audio embedding extraction and BERT for text embeddings, allowing the contributions of each modality to be thoroughly examined. The method demonstrates a significant improvement in summarization accuracy over text-only models. On the AMI [1] dataset, the proposed approach achieves ROUGE-1 (0.94), ROUGE-2 (0.93), ROUGE-L (0.94), and BLEU (0.88), outperforming state-of-the-art models such as BART and PEGASUS. Similarly, on the ICSI [2] dataset, it achieves ROUGE-1 (0.79), ROUGE-2 (0.75), and BLEU (0.47), showcasing its robustness across datasets. Furthermore, audio-based summarization achieves higher significance scores (94%) compared to text-based (89%), highlighting the additional context provided by auditory features. The findings indicate that audio features are crucial for capturing emotional tone, speaker intent, and conversational flow, which are often overlooked by text-based methods. By integrating both modalities, this framework produces more contextually rich and accurate summaries. This research sets a new benchmark for dialogue summarization by demonstrating that incorporating auditory cues leads to a deeper and more comprehensive understanding of dialogues, with applications ranging from virtual meetings to educational content summarization. ### Efficient Supernet Training with Orthogonal Softmax for Scalable ASR Model Compression ⚑ DE - **DOI:** [10.1109/icassp49660.2025.10890173](https://doi.org/10.1109/icassp49660.2025.10890173) - **Metadaten:** Erschienen: 2025-03-12 · S. 1-5 · OpenAlex seit 2025-10-10 - **DE-Institutionen:** RWTH Aachen University - **Autoren:** - Jingjing Xu — RWTH Aachen University (DE) - Eugen Beck — AppTek GmbH,Germany - Zijian Győző Yang — RWTH Aachen University (DE) - Ralf Schlüter — RWTH Aachen University (DE) - **Topics:** Speech Recognition and Synthesis; Advanced Neural Network Applications; Speech and Audio Processing - **Keywords:** model compression; speech recognition; CTC - **Abstract:** ASR systems are deployed across diverse environments, each with specific hardware constraints. We use supernet training to jointly train multiple encoders of varying sizes, enabling dynamic model size adjustment to fit hardware constraints without redundant training. Moreover, we introduce a novel method called OrthoSoftmax, which applies multiple orthogonal softmax functions to efficiently identify optimal subnets within the supernet, avoiding resource-intensive search. This approach also enables more flexible and precise subnet selection by allowing selection based on various criteria and levels of granularity. Our results with CTC on Librispeech and TED-LIUM-v2 show that FLOPs-aware component-wise selection achieves the best overall performance. With the same number of training updates from one single job, WERs for all model sizes are comparable to or slightly better than those of individually trained models. Furthermore, we analyze patterns in the selected components and reveal interesting insights.