Papers

Neural audio representation learning has enabled a range of content-oriented applications, but the resulting features remain limited for tasks involving audio processing. Furthermore, it is not obvious what processing-aware representations should capture: the processing itself, abstracted away from source content, or the processed audio that retains it. Existing approaches implicitly commit to one or the other and also differ in their models, data, and evaluation, obscuring which design choices drive their behavior. We address both questions within a unified framework of three objectives: processing consistency, description alignment, and equivariance via forward prediction. We compare all combinations of the objectives under a controlled setup and reveal their relative strengths and interactions. Our framework produces both a transformation embedding and a processed-audio embedding, and we find that the two play complementary roles: distance-based tasks favor the former, while probe-based tasks favor the latter. Combined with improvements in network architecture and training pipeline, our representations outperform prior baselines across retrieval, probe-based evaluation, and style transfer.

Prosody plays an important role in speech translation, conveying information such as emphasis, emotion, and intent beyond lexical content. However, despite recent progress in expressive speech-to-speech translation (S2ST), little is known about how prosodic patterns are similar/different across languages. Understanding these cross-lingual similarities and differences is crucial for effectively incorporating prosody into expressive S2ST systems. In this work, we present the first fine-grained cross-lingual analysis of prosody using multilingual dubbing data across English-German, English-Spanish, and English-French language pairs. We analyze the similarity of pitch, energy, and temporal feature patterns between source and target speech and investigate the linguistic and alignment-related factors affecting this similarity. Our analysis reveals inherent cross-lingual correlations in prosodic structure between certain languages. The findings provide important insights into the transferability of prosody across languages and offer empirical guidance for future expressive speech-to-speech translation systems.

Speech and audio LLMs are evaluated by comparing waveform predictions with predictions from an automatic speech recognition (ASR) transcript. For fixed closed-set tasks, this conflates acoustic evidence with the need to invoke a generative audio model. We estimate incremental call value with matched selectors sharing pre-call evidence. Each policy may retain the transcript label, use a local encoder, or invoke a generative model; matched control removes generative actions but preserves pre-call evidence and development selection. On VocalSound, transcript-only accuracy is 0.296, while supervised CLAP and WavLM controls reach 0.850 and 0.854 without calls. Full selector reaches 0.925 at 12.5% calls versus 0.921 for matched No-call selector (difference 0.004; 95% CI [-0.025, 0.033]). Thus, results do not show a call gain after transcript and encoder evidence are available. Relevant quantity is incremental accuracy from allowing calls, not the waveform-transcript gap.

Speech language models (speech LLMs) can generate plausible outputs from audio that contains no usable speech evidence. We study this failure as a pre-generation support-estimation problem and present SURE-Voice, a training-free front end that decides whether an audio prompt contains intelligible speech evidence before calling a speech LLM. We build SURE-Challenge with a 640-example SURE-Core split and a 1,920-example SURE-Extended split derived from 120 LibriSpeech source utterances. Using one fixed operating point, an energy screen plus Whisper token confidence raises unsupported accuracy on the held-out Extended test from 0.000–0.133 to 0.919 for six non-degenerate speech LLM backbones, while supported accuracy remains 0.919–0.970 and downstream calls fall from 480 to 287. A 500-clip ESC-50 sanity set shows the same pattern on real environmental audio, with vocal non-speech as a residual failure mode. An overlap diagnostic shows that source attribution remains separate from speech-evidence filtering. The evidence supports a controlled benchmark baseline and a deployment-oriented analysis; it does not establish universal robustness to semantic answerability, gain variation or natural conversations.

The relationship between subjective test vote variance and vote mean (or MOS) is well-studied, and the mathematically admissible vote variance region has been previously defined. We propose a reduced admissible variance region called the Unimodal Variance Region (UVR) that better describes real subjective rating behavior of multimedia. Further, subjective vote variance is often modeled as parabolic. We explain that, in practice, the parabolic model often violates the admissible region in the variance vs. MOS plane and we propose alternatives that respect the admissible region. We also present a parametrized random process to model votes that mixes voting processes and produces a realistic range of vote variances within the UVR at any desired MOS. This process was inspired by and comports with voting behavior that is observed in many subjective tests. By modeling vote variance from a subjective experiment, this vote model offers additional interpretable insights into voting behavior observed in a given experiment. We present example results from 16 datasets spanning speech, image, and video subjective quality experiments.

We present FDverb, a frequency-domain artificial reverberator, based on the idea of a vocoder with a noise carrier signal. Using a short-time Fourier transform (STFT) for analysis and synthesis, FDverb generates late reverberation by weighting spectral noise components with envelopes. We extend FDverb with early reflections, nonlinear decay, and pitch shifting. These extensions enable creative sound-design applications. We provide FDverb as an open-source DAW plug-in, using the JUCE framework.

Models mapping acoustic properties onto affective response underpin applications from music recommendation to sound design, yet are evaluated almost entirely within the corpus they were fitted on. When one fails outside it, the standard response – more data, or a larger model – assumes every failure is a shortage of resources. We show it is not, and that the alternative calls for the opposite remedy. Using four corpora of rated sound, four pretrained representations and three corpora of physiological recording, we pushed one mapping across four boundaries an application must cross: to new material, to edited audio, to a sensor in place of a self-report, and to an individual listener. At each we report the ceiling the target permits, the fraction surviving the crossing, and the price in target-side observations of closing the gap. Within a corpus, prediction reaches 84% of the ceiling set by inter-listener agreement. A same-domain corpus swap costs a fifth of that, and a hundred target labels return two-thirds of the loss. Crossing between music and environmental sound costs four-fifths to all of it, and four pretrained representations recover none of it. Against physiological response no information source we constructed exceeds a third of the attainable ceiling. "The model does not generalise" is therefore two diagnoses, not one, with mutually exclusive remedies; treating the second as the first is the more expensive mistake.

Oscillator synchronization is a widely used sound-synthesis technique, but straightforward digital implementations suffer from aliasing artifacts. This paper presents an alias-free method for digital emulation of oscillator synchronization of arbitrary periodic waveforms based on additive synthesis. Starting from a finite set of Fourier-series coefficients representing a bandlimited free-running waveform, we derive linear spectral-resampling transforms that map these coefficients to those of the bandlimited synchronized waveform. Beyond conventional hard synchronization, the proposed approach also supports two additional soft-synchronization modes. To address the high computational complexity of the proposed method, we introduce HASY, a 6 mm^2 application-specific integrated circuit (ASIC) fabricated in 65 nm CMOS technology. HASY generates one 96 kHz, 24 bit alias-free synchronized waveform with up to 512 harmonics and computes the spectral-resampling transform within only five audio-sample periods.

In electric guitars, the vibrations of the strings are typically sensed by coils of wire combined with a magnet, called pickups. The pickups and their position along the strings contribute strongly to the instrument's sound. Most guitars feature one to three pickups, each spanning across all strings with fixed positions and generating a single mono output. The work of this Master's Thesis at ETH Zürich introduces a new pickup system called PolyMap, which senses each string individually and at multiple locations. The system is demonstrated with a custom-made eight-string guitar that contains eight pickups per string for a total of 64 pickups. The signals from these 64 pickups are individually digitized inside the guitar and transmitted over a multichannel audio digital interface (MADI), a low-latency digital audio interface, to a computer for further processing. PolyMap enables high-resolution sensing of an electric guitar's strings and enables extensive post-processing capabilities for musicians, audio engineers, and researchers. To the best of our knowledge, this is the first polyphonic guitar pickup system with such a complete feature set.

Understanding spoken dialogue requires joint reasoning over lexical content and paralinguistic acoustic signals such as emotion and conversational intent. However, existing evaluations often allow shortcuts based on transcripts or single-modality solutions, obscuring whether models genuinely ground predictions in speech. We formalize this failure mode as cross-modal disagreement, where transcripts suggest plausible but incorrect surface interpretations while acoustic cues such as prosody or speaking style support different answers. We develop a scalable framework that identifies text-biased surface interpretations and converts disagreement regions into conflict QA examples. We also include consistent cases where transcript-based and speech-grounded interpretations agree, enabling evaluation beyond adversarial audio dependence. This results in ContraTalk, a controlled benchmark containing 501 questions across five discourse dimensions: interaction behavior, emotion state, dialogue act, social stance, and conversational intent. We further develop an agentic-style reasoning framework that converts speech into an Audio Twin, a text-readable representation of localized acoustic cues that exposes acoustic evidence to the reasoning model. Experiments show that strong text-only LLMs exceed 90% accuracy in consistent cases but drop to 33-48% in conflict cases. Direct AudioLLMs provide only partial grounding, still selecting the transcript-biased trap in roughly 30-40% of conflict cases. Our Audio Twin framework improves conflict-case accuracy while reducing trap selection, but its consistent-case behavior remains backbone-dependent. These results identify transcript-based shortcuts as an important failure mode in spoken dialogue understanding and show that explicit acoustic evidence aggregation provides a more controllable interface for diagnosing and improving speech-grounded reasoning.

In many low-resource settings, even just eliciting speech for data collection is difficult. One promising approach has been to ask speakers to describe images. But how do we build models from such visually grounded speech data? Given a dataset of images with Hindi spoken captions, we consider how we can map a written English keyword to spoken realisations of that word in Hindi. Previous work trained end-to-end multimodal neural models. Instead, we explore a simpler alignment-based approach built on self-supervised speech representations. Written English tags are automatically obtained from images using off-the-shelf image captioning systems. Hindi utterances associated with the same keyword are then aligned (using self-supervised features), and alignment evidence is aggregated to identify recurring speech segments corresponding to the target word. Experiments evaluating keyword spotting and localization show that our alignment-based approach outperforms a previous attention-based neural model. We also show the benefit of incorporating negative examples during alignment. Our work demonstrates that cross-lingual word-to-speech mappings can be learned directly from visual grounding without transcriptions or explicit model training.

Omnimodal large language models (Omni-LLMs) integrate audio, video, and text, yet remain vulnerable to cross-modal hallucinations, where one modality improperly influences predictions about another. Existing training-free decoders modulate modality influence through perturbation or relevance weighting, but do not assess predictive compatibility within the joint audio-visual branch. Because joint-branch discrepancies may indicate either harmful interference or useful complementarity, reliable intervention requires assessing both discrepancy magnitude and actionability. To this end, we propose Conflict-Aware Decoding (CAD), a training-free framework comprising Potential Conflict Magnitude Estimation (PCME) and Conflict Actionability Assessment (CAA). PCME quantifies potential conflict using audio-video disagreement and the deviation of the joint prediction from a relevance-weighted unimodal reference. CAA then applies Dempster-Shafer reliability discounting to task-space answer relations, using query relevance and answer decisiveness to determine whether intervention is warranted. When an actionable conflict is identified, CAD selectively reallocates decoding weight from the joint branch to the unimodal branches. Experiments on CMM, AVHBench, WorldSense, and VideoMME show that CAD consistently outperforms the base decoder and competitive training-free methods across multiple audio-visual backbones. On Qwen2.5-Omni-7B, CAD improves overall accuracy by 14.1 and 8.0 percentage points on CMM and AVHBench, respectively, without model retraining.

General audio comprehension now covers speech, sound, and music over durations from seconds to hours, driven by large audio-language models (LALMs) that are increasingly omni-modal. Yet the benchmarks that test them still rely on clips of seconds, where scores saturate and models converge; recent long-form efforts extend duration but evaluate long audio much as short clips are. We introduce LongAudioSpan, a benchmark that spans both duration and depth: it pairs audio from 10 minutes to over 2 hours with 3,240 questions across three cognitive levels, namely perception, understanding, and reasoning. Two paths supply the questions, differing in how question content is sourced and how ground truth is obtained. Native QA extracts questions from the audio's content, posing each as a multiple-choice item and an open-ended one graded by detailed rubrics. Anchor QA instead injects ground truth, planting acoustic anchors into the audio and building a perception-to-reasoning chain scored only to the first error. A fully automated pipeline constructs every item through structured captioning, QA generation, and adversarial critic feedback. Evaluating 12 LALMs on LongAudioSpan, we find the hard part comes before reasoning: distilling a few relevant facts from a long, redundant signal. This difficulty grows with audio length and falls hardest on perception, especially temporal grounding. LongAudioSpan is available at https://huggingface.co/datasets/holvan/LongAudioSpan.

Recently, neural directional filtering (NDF) enables reconstruction of a virtual directional microphone (VDM) with a desired directivity pattern, accurately rendering multi-source scenes by preserving spatial cues. In strongly reverberant environments, spatial cues become perceptually difficult to distinguish, limiting NDF-based spatial sound capture. This paper addresses this limitation with three contributions: First, we propose a neural dereverberation and directional filtering (NDDF) approach to reconstruct dereverberated VDM signals. Second, NDDF is implemented with discriminatively trained and generative adversarial network (GAN)-based models, compared with cascaded dereverberation and directional-filtering baselines. Experimental results indicate that the NDDF consistently surpasses the cascaded baselines. Additionally, the GAN-based NDDF outperforms the discriminative variant when addressing a high-order VDM target. Third, we introduce a method for directivity pattern estimation that relies solely on the input and output signals. This method is suitable for signal-mapping-based spatial filtering, which synthesizes the output signal directly without explicit filtering or masking.

Large language model (LLM)-based audio-visual speech recognition (AVSR) systems are robust under noise. Contrastive decoding (CD), originally introduced to stabilize LLM generation by contrasting a weaker model against a stronger one at inference time, adjusts predictions without additional training. In this work, we apply CD to AVSR by contrasting audio-only conditioning with full audio-visual conditioning within the same underlying model. However, using a fixed contrastive strength introduces a trade-off across noise levels: stronger intervention helps under severe noise but may over-correct reliable predictions in clean conditions. We propose reliability-aware scaling of CD for AVSR. Instead of using a fixed strength, we adaptively modulate the contrastive influence at each token based on reliability signals derived from attention dynamics and inter-model predictive divergence. Experiments on LRS3 show consistent improvements across clean and low-SNR conditions.

Conversational systems, such as duplex speech language models (SLMs), still lack a streaming, accurate, and empathetic memory system as their soul. We introduce VoiceMem, a simple memory architecture with a parallel informational left brain, an emotional right brain, and streaming memory I/O mechanisms. We further build a complete pipeline for memory-aware SLM training, long-horizon evaluation, and decoupled deployment with interchangeable memory backends. Experiments and real-world deployment show three advantages: i) Accuracy: under top-5 retrieval, the left brain outperforms classical systems such as Mem0 at top-200 by nearly 30 points; ii) Emotional & Personal: the right brain, with short- and long-horizon affective attribution and dual-node persona modeling, achieves state-of-the-art performance across three persona benchmarks and improves the aggregate score by 4.29 points over the previous best system; and iii) Real-Time & Cheap: VoiceMem completes retrieval in 134 ms, well within standard VAD latency, adding no extra conversational delay while maintaining high accuracy and low cost. These results show that VoiceMem provides a practical memory foundation for real-time, personalized, and emotionally aware speech interaction.

Air traffic control procedures are executed through spoken exchanges between controllers and pilots. These interactions are essential to the safety of air transportation: failures in their execution can create severe operational hazards, as evidenced by past fatal accidents. Assessing whether an instruction has been followed requires relating what was said to the aircraft concerned, its state, and the obligations that pilots must meet. We present a runtime verification framework that monitors such procedures by checking controller-pilot exchanges, surveillance data, and onboard observations. The framework parses radio communications into events linked to the entities they concern and merges them with surveillance and onboard observations into a time-stamped trace. The ICAO-derived obligations as formalized as temporal formulas with explicit time bounds and evaluated over execution traces. Every violation is reported along with the breached obligations and the observations that support the verdict. With real traffic, the complete pipeline reaches an F1 of 0.85 against blind human-annotated violations; in 1,495 synthetic situations derived from two public corpora, the monitor logic returns the expected verdict in every case. In two historical accidents reconstructed from official investigation reports, the monitor identifies the same procedural deviations documented by the investigators.

Cough acoustics are promising for non-invasive tuberculosis (TB) screening, yet whether machine learning (ML) models capture disease-related acoustics or artifacts of data collection remains unresolved. We evaluated the cross-dataset generalizability of classical ML and deep learning (DL) cough-based TB classifiers across three independent datasets. Despite moderate within-dataset performance (ROC-AUC up to $0.755 \pm 0.056$), both pipelines fail to generalize, with external performance frequently below 0.6, indicating a possible limitation of the data. We further observed audio representations are organized by recording device and dataset rather than TB status, predicted TB probability tracks country-level prevalence in CODA, and device mismatch degrades transfer while device-diverse training improves it. Additionally, a clinical-variable baseline generalizes more consistently (ROC-AUC $0.655 - 0.711$), indicating acquisition-specific variability is a stronger driver of poor generalizability than population shift. High within-dataset performance is not enough. External validation is essential before cough-based TB models are clinically ready.

Federated Learning (FL) enables distributed training of machine learning models while preserving data privacy. However, FL struggles with heterogeneous, non-IID client data distributions, resulting in sub-optimal and biased global models. In this paper, we propose pFedMARL, a novel approach leveraging Multi-Agent Reinforcement Learning (MARL) with Twin Delayed Deep Deterministic Policy Gradient (TD3) to dynamically adapt aggregation strategies in FL settings. Our method employs a server-side agent adjusting client contributions to optimize global model robustness and client-side agents balancing global and local updates to personalize models effectively without pre-training. We demonstrate superior performance of pFedMARL for training a semi-supervised audio spectrogram transformer, matching or outperforming FedAvg, Ditto, and local training approaches across multiple non-IID scenarios and in the presence of adversarial clients. Our results indicate that pFedMARL actively improves accuracy, robustness, and fairness, making it suitable for real-world deployments.

In recent years, many efforts have been made to supersede classical acoustic echo control (AEC) algorithms with more powerful machine-learned approaches. While surpassing the performance of well-established adaptive filters is very much possible, a remaining challenge is computational complexity. Popular architectures, such as convolutional recurrent networks (CRNs), are by multiple orders of magnitude computationally more expensive than classical signal processing solutions. Scaling down such models is usually straight-forward, but it comes at the cost of a notably reduced performance. We show - to the author's knowledge for the first time in AEC - how these performance drops can be successfully alleviated to a large degree by employing an effective knowledge distillation (KD) process, enabling more potent efficient AEC. Our proposed CGGN16 student AEC models show significantly less near-end speech distortion at only 2% of its teacher's computational complexity, surpass the overall performance of a six times more complex model trained on ground-truth labels, and outperform other AEC-focused architectures from recent literature.

This paper proposes acoustic echo control based on sound object identification for suppressing acoustic echo and howling in conferencing environments with complicated acoustic paths, where multiple hands-free terminals coexist in the same room. Conventional acoustic echo cancellers target fixed intra-device echo paths; however, unintended paths, for example, those formed via inter-terminal communication, are difficult to control and can lead to howling. Instead of estimating echo paths, the proposed approach identifies sound objects and keeps channels muted by default, allowing pass/playback only when the signal is judged not dentical to recently observed objects. This breaks echo loops caused by repeated reproduction of the same sound object and can be viewed as an extension of classical voice switching toward conditional half-duplex operation. Technical challenges and connections to related techniques are discussed, and a simulation shows howling suppression together with a trade-off against speech quality.

Spatial audio vocoders are able to convert mel-spectrograms produced by generative models into spatial audio waveforms. Most neural vocoders are designed for monaural audio, and direct extensions to spatial audio can degrade spatial quality by ignoring inter-channel cues. We present CSAVocoder, a causal GAN-based spatial audio vocoder that jointly optimizes waveform fidelity and spatial rendering. Our framework introduces a Spatial Adaptor that fuses multi-channel mel-spectrograms with dynamic source-listener pose information, together with a spatial consistency discriminator that supervises inter-channel cues. To meet real-time requirements, we design a strictly causal, stateful generator that supports efficient streaming inference with constant memory overhead. Experiments on large-scale spatial audio datasets show that CSAVocoder improves spatial fidelity at competitive audio quality and real-time performance.

Mandarin homophones remain a key challenge to improving automatic speech recognition (ASR) accuracy due to the amount of potential homophones. Mandarin speakers use this feature casually to convey emotions such as humour. Recent homophone-aware ASR studies have improved recognition accuracy, but intentional homophone twists in speech remain underexplored. In this paper, we identify patterns of homophone-based rhetorical wordplay in Mandarin, referred to as HumourPhone, and propose an ASR Adapter for homophone and HumourPhone recovery. Experimental results show that the proposed approach improves the recall of recognising HumourPhone by over 5% and achieves a 4.35% drop in target-span character error rate for homophone correction compared to baseline. These results highlight the need for homophone-aware modelling of lexical ambiguity and rhetorical wordplay in Mandarin speech recognition.

Speakers in natural conversation take turns speaking and listening, deciding in real time when to take, hold, or yield the floor. However, turn-taking evaluation remains limited due to the lack of a consistent, linguistically grounded evaluation protocol and hand-annotated data covering diverse conversation types. To address this, we present TurnBench, a multi-domain benchmark that pairs a 30-hour, hand-labeled corpus of dyadic human conversation with a standardized evaluation protocol for end-of-turn and interruption detection. We set conversation type as a controllable experimental variable, covering six distinct interaction styles, and triple-annotate each conversation. Benchmarking 14 heterogeneous turn-taking systems, we find end-of-turn recall stable across types, while interruption false positives are strongly type-dependent and concentrated in backchannel-dense interaction styles. Although in smooth floor transfers human listeners begin speaking a median 151 ms before the current turn ends, no current system performs equivalently without incurring excessive false positives. We release our corpus, a 104-hour training set, and a public leaderboard with an interactive dataset viewer at https://turnbench.sesame.com.

We propose $\textit{recursive encoder and decoder}$ (RED) for building a single deep neural network (DNN) model that can separate multi-speaker mixtures containing unknown numbers of speakers and variable numbers of microphones arranged in an unknown geometry, a task that has not been studied yet. The decoder of RED recursively detects whether there are active speakers left and separates one speaker at a time. It is designed to be trained in an end-to-end fashion to improve separation performance. The encoder of RED recursively encodes each microphone channel of the input mixture, sequentially incorporating spatial cues. Combining both, the DNN can be trained to separate mixtures not only with unknown numbers of speakers but also with variable numbers of microphones, achieving state-of-the-art performance on multiple public datasets.

Deep neural network based deepfake detection systems have achieved high levels of accuracy on benchmark datasets and competitions. However, most models lack interpretability. It is challenging to extract reasoning from the network that can convince the human evaluator to trust the decision. Humans often rely on acoustic cues like unnatural pitch jitter, robotic intonation, acoustic artifacts, and unnatural sounding fricatives to judge the quality of the synthetic audio. This study explores the role played by the voiced and unvoiced regions of speech in discriminating synthetic from bonafide speech. A measure of signal periodicity is used to analyze speech into voiced and unvoiced components. Then, the graph attention based AASIST detection system is trained independently on each component. This work compares the accuracy of deepfake detection system using voiced and unvoiced components and analyzes the results on the MLAAD dataset. Our results show that unvoiced regions are particularly more effective in distinguishing synthetic (deepfake) speech from bonafide, and achieves an equal error rate of 6.62%. When combined with voice regions through score-level fusion, the overall performance improves further, yielding a 5.82% EER, a relative improvement of 49% over the baseline system that uses the full audio.

Spatial audio is a key component of immersive $360^\circ$ media, yet high-quality spatial capture remains limited in real-world speech-dominant scenes. We study visually guided First-Order Ambisonics (FOA) speech spatialization in the wild: given aligned $360^\circ$ video and an omnidirectional audio track, we recover the missing directional FOA components. To support this task, we introduce YT-SPEECH, a speech-oriented $360^\circ$ video-FOA dataset curated from YouTube. We propose a two-stage Localizer-Renderer framework, where an audio-visual segmentation backbone provides frame-wise spatial heatmaps and a conditional complex-domain U-Net reconstructs directional FOA signals from the omnidirectional channel. A confidence-based gating strategy stabilizes conditioning under ambiguous acoustic conditions. Experiments show improved reconstruction fidelity, spatial accuracy, and perceptual speech quality relative to ablated variants and prior approaches.

Spatial audio enhances user immersion by reproducing 3D sound fields, with Ambisonics being a widely adopted representation. While Ambisonics is theoretically independent of the recording setup, practical microphone arrays introduce hardware-dependent encoding artifacts. Moreover, existing data-driven solutions lack flexibility, as they are typically restricted to fixed array geometries. To overcome these limitations, we propose ADEPS, a generative framework that explicitly embeds the physical acquisition model into the inference process. By leveraging this formulation, ADEPS effectively compensates for array-specific distortions while enabling zero-shot encoding across arbitrary array topologies. We train the underlying generative prior in an unsupervised manner solely on target Ambisonic representations. Extensive evaluations across diverse simulated and real microphone arrays demonstrate that ADEPS consistently outperforms both traditional linear and parametric baselines in spatial fidelity and spectral quality.

With the advent of Large Language Models and its instruction following capabilities a promising application is the task of summarization. Within this domain of task the extractive sub-task of clinical protocolling has emerged as a topic of particular interest as it can significantly reduce the downtime and protocolling burden of health-care workers thus enabling them to focus on their core work helping humans. A further step towards automation is the direct generation of clinical notes from speech without intermediate transcripts, reducing processing time while preserving information such as coughing or other paralinguistic cues that may be lost in transcript-based systems. To this end, we present KIT's submission to this years BeTraC challenge in the lightweight track. Our main contribution is a scalable data augmentation pipeline that unifies heterogeneous medical dialogue datasets through synthetic speech generation and automatically generated SOAP supervision, enabling robust adaptation of a speech foundation model for end-to-end speech-to-SOAP generation.

Non-verbal vocalizations (NVs), such as laughter, coughs, and sighs, are essential for expressive TTS, but the effectiveness of preference optimization for NV generation remains poorly understood. We systematically study preference optimization for NV-capable TTS, focusing on preference signals, preference-pair construction, and DPO-based optimization objectives. We formulate an NV-aware character error rate (NV-CER) by treating NV tags as distinct output symbols and computing a weighted pinyin-based CER over both verbal and non-verbal content, enabling controllable optimization of NV realization without modifying the underlying optimization algorithm. Experiments on Emilia-NV and the augmented NV-Bench covering 18 NV types reveal how different design choices affect NV realization and lexical fidelity, and establish an effective setup using standard DPO. Objective, LLM-based, and human evaluations provide converging evidence for our findings, offering practical insights into NV-aware post-training for expressive TTS.