Ting Dang

Di Zhu, Yu Yvonne Wu, Hong Jia et al.

Wearable devices enable continuous monitoring of physiological signals such as ECG and PPG, but existing mHealth systems are largely limited to task-specific prediction pipelines or reactive question answering over static summaries. They lack the ability to support temporal reasoning, persistent physiological context, and proactive monitoring over long-term signal streams. We propose VitalAgent, a tool-augmented agentic framework for ECG/PPG-based mHealth that supports both reactive question answering and proactive monitoring. VitalAgent is built on a longitudinal physiological memory and a tool-augmented reasoning interface that enables dynamic computation over raw signals. We further introduce VitalBench, a longitudinal physiological monitoring benchmark dataset comprising 1,862 QA pairs for reactive question answering and 90.2 hours of continuous ECG/PPG recordings for proactive monitoring, covering cardiac, physical activity, and stress-related tasks. Experiments demonstrate that VitalAgent achieves over 30% improvement over prompt-based and ReAct baselines in reactive evaluation and supports proactive alert monitoring over long-term physiological signals, highlighting the importance of dynamic tool use and long-term physiological monitoring.

Kai Bian, Xucheng Guo, Bin Chen et al.

Evaluations of dental vision-language models remain fragmented across datasets, task definitions and metrics, and often ignore their computational cost. This limits their widespread deployment for dental screening outside specialist centres, where timely inference, limited hardware, and local handling of patient images are vital for practical, privacy-preserving clinical prescreening. Here we present Pocket-Dentist, an efficiency-aware benchmark for dental multimodal question answering that brings together three datasets spanning approximately 1,159 patients, five task types and seven metrics. Across typical 14 VLMs, our results reveals an interesting observation: compact VLMs (e.g., 2B-parameter models) outperform larger VLMs in accuracy while requiring substantially lower computational costs in dental image understanding. Deployed locally on an iPhone 17 Pro, our finetuned compact VLM Pocket-Dentist-2B processed each sample in 4.31 s, reducing latency by 4.9-fold and memory use by 2.3-fold compared with a 7B baseline.

Tianyi Zhang, Xiangyuan Xue, Lingyan Ruan et al.

Mental health conditions affect hundreds of millions globally, yet early detection remains limited. While large language models (LLMs) have shown promise in mental health applications, their size and computational demands hinder practical deployment. Small language models (SLMs) offer a lightweight alternative, but their use for social media--based mental health prediction remains largely underexplored. In this study, we introduce Menta, the first optimized SLM fine-tuned specifically for multi-task mental health prediction from social media data. Menta is jointly trained across six classification tasks using a LoRA-based framework, a cross-dataset strategy, and a balanced accuracy--oriented loss. Evaluated against nine state-of-the-art SLM baselines, Menta achieves an average improvement of 15.2\% across tasks covering depression, stress, and suicidality compared with the best-performing non--fine-tuned SLMs. It also achieves higher accuracy on depression and stress classification tasks compared to 13B-parameter LLMs, while being approximately 3.25x smaller. Moreover, we demonstrate real-time, on-device deployment of Menta on an iPhone 15 Pro Max, requiring only approximately 3GB RAM. Supported by a comprehensive benchmark against existing SLMs and LLMs, Menta highlights the potential for scalable, privacy-preserving mental health monitoring. Code is available at: https://xxue752-nz.github.io/menta-project/

Han Yin, Yang Xiao, Younghoo Kwon et al.

Large audio language models (LALMs) are a class of foundation models for audio understanding. Existing LALMs tend to degrade significantly in real-world noisy acoustic conditions where speech and non-speech sounds interfere. While noise-aware fine-tuning can improve robustness, it requires task-specific noisy data and expensive retraining, limiting scalability. To address this issue, we propose Focus-Then-Listen (FTL), a plug-and-play audio enhancer that improves LALMs' noise robustness. Specifically, FTL first separates the input waveform into speech and non-speech, and a modality router is applied to predict the target audio modality (e.g., speech) based on the user's instruction. Finally, a modality-aware fusion block generates a task-adaptive enhanced signal for improved downstream perception and reasoning. Experiments across multiple LALMs and tasks show that FTL improves performance across different noise levels without fine-tuning on LALMs.

Xin Wang, Ting Dang, Vassilis Kostakos et al.

Healthcare monitoring is crucial for early detection, timely intervention, and the ongoing management of health conditions, ultimately improving individuals' quality of life. Recent research shows that Large Language Models (LLMs) have demonstrated impressive performance in supporting healthcare tasks. However, existing LLM-based healthcare solutions typically rely on cloud-based systems, which raise privacy concerns and increase the risk of personal information leakage. As a result, there is growing interest in running these models locally on devices like mobile phones and wearables to protect users' privacy. Small Language Models (SLMs) are potential candidates to solve privacy and computational issues, as they are more efficient and better suited for local deployment. However, the performance of SLMs in healthcare domains has not yet been investigated. This paper examines the capability of SLMs to accurately analyze health data, such as steps, calories, sleep minutes, and other vital statistics, to assess an individual's health status. Our results show that, TinyLlama, which has 1.1 billion parameters, utilizes 4.31 GB memory, and has 0.48s latency, showing the best performance compared other four state-of-the-art (SOTA) SLMs on various healthcare applications. Our results indicate that SLMs could potentially be deployed on wearable or mobile devices for real-time health monitoring, providing a practical solution for efficient and privacy-preserving healthcare.

Kaiwen Luo, Zhenhong Zhou, Leo Wang et al.

The foundational capabilities established by Large Language Models (LLMs) have paved the way for Multimodal Large Language Models (MLLMs), within which Large Audio Language Models (LALMs) are essential for realizing universal auditory intelligence. Despite their remarkable performance, the escalation of LALMs' capabilities has significantly outpaced the development of systemic frameworks to ensure their trustworthiness. This survey provides a comprehensive investigation into the endogenous mechanisms of LALMs, detailing the architectural innovations and alignment algorithms that facilitate emergent reasoning. Specifically, we analyze how the transition to unified end-to-end frameworks and the integration of continuous acoustic signals inherently expand the attack surface. To rigorously evaluate the risks within these paradigms, we establish a comprehensive taxonomy of trustworthiness, categorizing critical vulnerabilities such as cross-modal jailbreaking, latent acoustic backdoors, and biometric privacy leakage. We review the state-of-the-art through six analytical pillars: hallucination, robustness, safety, privacy, fairness, and authentication. The profound imbalance between a mature offensive landscape and underdeveloped defenses further validates the critical trustworthiness gaps and multidimensional risks facing audio-centric intelligence. Finally, we propose a strategic roadmap advocating for "Defense-in-Depth" architectures, causal auditory world modeling, and intrinsic representation engineering to bridge the gap between empirical performance and intrinsically trustworthy audio intelligence. Our project has been uploaded to GitHub https://github.com/Kwwwww74/Awesome-Trustworthy-AudioLLMs.

Jingyao Wu, Ting Dang, Vidhyasaharan Sethu et al.

There has been a significant focus on modelling emotion ambiguity in recent years, with advancements made in representing emotions as distributions to capture ambiguity. However, there has been comparatively less effort devoted to the consideration of temporal dependencies in emotion distributions which encodes ambiguity in perceived emotions that evolve smoothly over time. Recognizing the benefits of using constrained dynamical neural ordinary differential equations (CD-NODE) to model time series as dynamic processes, we propose an ambiguity-aware dual-constrained Neural ODE approach to model the dynamics of emotion distributions on arousal and valence. In our approach, we utilize ODEs parameterised by neural networks to estimate the distribution parameters, and we integrate additional constraints to restrict the range of the system outputs to ensure the validity of predicted distributions. We evaluated our proposed system on the publicly available RECOLA dataset and observed very promising performance across a range of evaluation metrics.

Xin Hong, Yuan Gong, Vidhyasaharan Sethu et al.

Recent advancements in Large Language Models (LLMs) have demonstrated great success in many Natural Language Processing (NLP) tasks. In addition to their cognitive intelligence, exploring their capabilities in emotional intelligence is also crucial, as it enables more natural and empathetic conversational AI. Recent studies have shown LLMs' capability in recognizing emotions, but they often focus on single emotion labels and overlook the complex and ambiguous nature of human emotions. This study is the first to address this gap by exploring the potential of LLMs in recognizing ambiguous emotions, leveraging their strong generalization capabilities and in-context learning. We design zero-shot and few-shot prompting and incorporate past dialogue as context information for ambiguous emotion recognition. Experiments conducted using three datasets indicate significant potential for LLMs in recognizing ambiguous emotions, and highlight the substantial benefits of including context information. Furthermore, our findings indicate that LLMs demonstrate a high degree of effectiveness in recognizing less ambiguous emotions and exhibit potential for identifying more ambiguous emotions, paralleling human perceptual capabilities.

Yang Xiao, Siyi Wang, Han Yin et al.

Large audio language models (LALMs) process both speech and environmental acoustic cues, yet struggle to retain non-speech information across multi-turn interactions. The performance gap between semantic (speech) and acoustic (non-speech) understanding remains poorly understood, and the underlying mechanisms of representation and retrieval are still unclear. This work introduces EnvMem, a controlled multi-turn benchmark designed to study this gap and identify the root causes of failures at the representation (i.e., latent embeddings) and retrieval levels (i.e., attention allocation). We further conduct post-hoc interventions to probe representational structure and attention dynamics. Our results reveal representational trajectory drift as the key failure mode, while showing that attention allocation plays a limited role in explaining the observed degradation. Overall, we provide a systematic framework for analyzing and improving non-linguistic memory in long-context LALMs, shedding light on future data and training design for robust acoustic memory modeling.

Feixiang Zheng, Yu Wu, Cecilia Mascolo et al.

Time series data from the Intensive Care Unit (ICU) provides critical information for patient monitoring. While recent advancements in applying Large Language Models (LLMs) to time series modeling (TSM) have shown great promise, their effectiveness on the irregular ICU data, characterized by particularly high rates of missing values, remains largely unexplored. This work investigates two key components underlying the success of LLMs for TSM: the time series encoder and the multimodal alignment strategy. To this end, we establish a systematic testbed to evaluate their impact across various state-of-the-art LLM-based methods on benchmark ICU datasets against strong supervised and self-supervised baselines. Results reveal that the encoder design is more critical than the alignment strategy. Encoders that explicitly model irregularity achieve substantial performance gains, yielding an average AUPRC increase of $12.8\%$ over the vanilla Transformer. While less impactful, the alignment strategy is also noteworthy, with the best-performing semantically rich, fusion-based strategy achieving a modest $2.9\%$ improvement over cross-attention. However, LLM-based methods require at least 10$\times$ longer training than the best-performing irregular supervised models, while delivering only comparable performance. They also underperform in data-scarce few-shot learning settings. These findings highlight both the promise and current limitations of LLMs for irregular ICU time series. The code is available at https://github.com/mHealthUnimelb/LLMTS.

Zheng Nan, Ting Dang, Vidhyasaharan Sethu et al.

Connectionist temporal classification (CTC) is commonly adopted for sequence modeling tasks like speech recognition, where it is necessary to preserve order between the input and target sequences. However, CTC is only applied to deterministic sequence models, where the latent space is discontinuous and sparse, which in turn makes them less capable of handling data variability when compared to variational models. In this paper, we integrate CTC with a variational model and derive loss functions that can be used to train more generalizable sequence models that preserve order. Specifically, we derive two versions of the novel variational CTC based on two reasonable assumptions, the first being that the variational latent variables at each time step are conditionally independent; and the second being that these latent variables are Markovian. We show that both loss functions allow direct optimization of the variational lower bound for the model log-likelihood, and present computationally tractable forms for implementing them.

Yang Xiao, Siyi Wang, Eun-Jung Holden et al.

Speech and audio systems operate in inherently non-stationary environments, yet continual learning (CL) research in this domain, especially in the foundation model era, remains fragmented that fail to account for the coupled, geometry-sensitive nature of acoustic representations. Modern speech foundation models operate over highly entangled, continuous representations that jointly encode linguistic, speaker, and paralinguistic factors within a shared latent space. CL is therefore fundamentally about preserving and evolving shared representation structure rather than retaining isolated task knowledge. In this work, we revisit CL for speech from a representation-centered perspective, and introduce a new taxonomy that organizes CL according to how underlying representation geometry evolves under non-stationary acoustic conditions. We further identify key mismatches between current CL assumptions and speech foundation model behavior, and finally outline a set of open challenges and future research directions.

Trevor Adelson, Vidhyasaharan Sethu, Ting Dang

Deep learning dominates speech processing but relies on massive datasets, global backpropagation-guided weight updates, and produces entangled representations. Assembly Calculus (AC), which models sparse neuronal assemblies via Hebbian plasticity and winner-take-all competition, offers a biologically grounded alternative, yet prior work focused on discrete symbolic inputs. We introduce an AC-based speech processing framework that operates directly on continuous speech by combining three key contributions:(i) neural encoding that converts speech into assembly-compatible spike patterns using probabilistic mel binarisation and population-coded MFCCs; (ii) a multi-area architecture organising assemblies across hierarchical timescales and classes; and (iii) cross-area update schemes for downstream tasks. Applied to two core tasks of boundary detection and segment classification, our framework detects phone (F1=0.69) and word (F1=0.61) boundaries without any weight training, and achieves 47.5% and 45.1% accuracy on phone and command recognition. These results show that AC-based dynamical systems are a viable alternative to deep learning for speech processing.

Puyue Wang, Jiawei Hu, Yan Gao et al.

Humanoid robots can suffer significant performance drops under small changes in dynamics, task specifications, or environment setup. We propose HoRD, a two-stage learning framework for robust humanoid control under domain shift. First, we train a high-performance teacher policy via history-conditioned reinforcement learning, where the policy infers latent dynamics context from recent state--action trajectories to adapt online to diverse randomized dynamics. Second, we perform online distillation to transfer the teacher's robust control capabilities into a transformer-based student policy that operates on sparse root-relative 3D joint keypoint trajectories. By combining history-conditioned adaptation with online distillation, HoRD enables a single policy to adapt zero-shot to unseen domains without per-domain retraining. Extensive experiments show HoRD outperforms strong baselines in robustness and transfer, especially under unseen domains and external perturbations. Code and project page are available at https://tonywang-0517.github.io/hord/.

Zheng Nan, Ting Dang, Vidhyasaharan Sethu et al.

Relational thinking refers to the inherent ability of humans to form mental impressions about relations between sensory signals and prior knowledge, and subsequently incorporate them into their model of their world. Despite the crucial role relational thinking plays in human understanding of speech, it has yet to be leveraged in any artificial speech recognition systems. Recently, there have been some attempts to correct this oversight, but these have been limited to coarse utterance-level models that operate exclusively in the time domain. In an attempt to narrow the gap between artificial systems and human abilities, this paper presents a novel spectro-temporal relational thinking based acoustic modeling framework. Specifically, it first generates numerous probabilistic graphs to model the relationships among speech segments across both time and frequency domains. The relational information rooted in every pair of nodes within these graphs is then aggregated and embedded into latent representations that can be utilized by downstream tasks. Models built upon this framework outperform state-of-the-art systems with a 7.82\% improvement in phoneme recognition tasks over the TIMIT dataset. In-depth analyses further reveal that our proposed relational thinking modeling mainly improves the model's ability to recognize vowels, which are the most likely to be confused by phoneme recognizers.

Yu Yvonne Wu, Yuwei Zhang, Hyungjun Yoon et al.

Wearable foundation models (WFMs), trained on large volumes of data collected by affordable, always-on devices, have demonstrated strong performance on short-term, well-defined health monitoring tasks, including activity recognition, fitness tracking, and cardiovascular signal assessment. However, most existing WFMs primarily map short temporal windows to predefined labels via static encoders, emphasizing retrospective prediction rather than reasoning over evolving personal history, context, and future risk trajectories. As a result, they are poorly suited for modeling chronic, progressive, or episodic health conditions that unfold over weeks, months or years. Hence, we argue that WFMs must move beyond static encoders and be explicitly designed for longitudinal, anticipatory health reasoning. We identify three foundational shifts required to enable this transition: (1) Structurally rich data, which goes beyond isolated datasets or outcome-conditioned collection to integrated multimodal, long-term personal trajectories, and contextual metadata, ideally supported by open and interoperable data ecosystems; (2) Longitudinal-aware multimodal modeling, which prioritizes long-context inference, temporal abstraction, and personalization over cross-sectional or population-level prediction; and (3) Agentic inference systems, which move beyond static prediction to support planning, decision-making, and clinically grounded intervention under uncertainty. Together, these shifts reframe wearable health monitoring from retrospective signal interpretation toward continuous, anticipatory, and human-aligned health support.

Yan Gao, Massimo Roberto Scamarcia, Javier Fernandez-Marques et al.

Large Language Models (LLMs) have achieved state-of-the-art results across diverse domains, yet their development remains reliant on vast amounts of publicly available data, raising concerns about data scarcity and the lack of access to domain-specific, sensitive information. Federated Learning (FL) presents a compelling framework to address these challenges by enabling decentralized fine-tuning on pre-trained LLMs without sharing raw data. However, the compatibility and performance of pre-trained LLMs in FL settings remain largely under explored. We introduce the FlowerTune LLM Leaderboard, a first-of-its-kind benchmarking suite designed to evaluate federated fine-tuning of LLMs across four diverse domains: general NLP, finance, medical, and coding. Each domain includes federated instruction-tuning datasets and domain-specific evaluation metrics. Our results, obtained through a collaborative, open-source and community-driven approach, provide the first comprehensive comparison across 26 pre-trained LLMs with different aggregation and fine-tuning strategies under federated settings, offering actionable insights into model performance, resource constraints, and domain adaptation. This work lays the foundation for developing privacy-preserving, domain-specialized LLMs for real-world applications.

Haoguang Zhou, Siyi Wang, Jingyao Wu et al.

Modern speech systems increasingly use discretized self-supervised speech representations for compression and integration with token-based models, yet their impact on emotional information remains unclear. We study how residual vector quantization (RVQ) reshapes emotional information in discrete speech representations from both representation- and task-level perspectives. Our analysis shows that aggressive compression disproportionately degrades emotion, with uneven loss across emotion classes and model architectures. To address this, we introduce emotion-aware quantization using emotion-specific and emotion-biased codebooks, improving the preservation of both hard and soft emotion perception. We further propose Emo-Q, a lightweight routed quantization method that selects emotion-specialized codebooks, improving emotion recognition performance at lower bitrates. These results highlight the importance of emotion-aware discretization for robust affective speech processing.

Amaan Izhar, Nurul Japar, Norisma Idris et al.

Medical image reporting (MIR) aims to generate structured clinical descriptions from radiological images. Existing methods struggle with fine-grained feature extraction, multimodal alignment, and generalization across diverse imaging types, often relying on vanilla transformers and focusing primarily on chest X-rays. We propose MicarVLMoE, a vision-language mixture-of-experts model with gated cross-aligned fusion, designed to address these limitations. Our architecture includes: (i) a multiscale vision encoder (MSVE) for capturing anatomical details at varying resolutions, (ii) a multihead dual-branch latent attention (MDLA) module for vision-language alignment through latent bottleneck representations, and (iii) a modulated mixture-of-experts (MoE) decoder for adaptive expert specialization. We extend MIR to CT scans, retinal imaging, MRI scans, and gross pathology images, reporting state-of-the-art results on COVCTR, MMR, PGROSS, and ROCO datasets. Extensive experiments and ablations confirm improved clinical accuracy, cross-modal alignment, and model interpretability. Code is available at https://github.com/AI-14/micar-vl-moe.

Xiangyuan Xue, Jiajun Lu, Yan Gao et al.

Speech Emotion Captioning (SEC) leverages large audio-language models to generate rich, context-aware affective descriptions from speech. However, real-world deployment remains challenging due to the substantial computational demands on resource-constrained edge devices and the privacy risks of transmitting biometric audio. While smaller audio-language models enable efficient on-device SEC, their limited capacity often weakens subtle paralinguistic modeling and fine-grained affective grounding. We propose an edge-cloud collaborative framework based on Uncertainty-Guided Speculative Decoding (UGSD). A lightweight edge model drafts captions locally, and only high-uncertainty token blocks are selectively escalated to a stronger cloud verifier for validation. Experiments on the MER2024 benchmark demonstrate substantial BLEU improvements up to 62.7%. UGSD further achieves 1.4x lower latency and 8.5x higher token throughput compared to an edge-only model. These results empirically characterize the quality-efficiency-privacy trade-off in deployable SEC systems.

Sung Kyun Chung, Jiaheng Dong, Qiuchi Hu et al.

Large Audio-Language Models (LALMs) have shown strong performance in speech understanding, making speech a natural interface for accessing factual information. Yet they are trained on static corpora and may encode incorrect facts. Existing model editing methods localize and update facts in text-only LLMs, but do not account for continuous speech representations, or where knowledge is stored across acoustic or language modules, or their cross-modal module. We construct the first audio benchmark for knowledge localization and editing in LALMs and propose a speech-driven locate-then-edit framework. First, we use speech-aware causal tracing to localize layers and modules that support factual retrieval and then apply editing at identified sites. Experiments show that factual knowledge is jointly encoded in audio and text modules, and that audio editing yields more effective updates than text editing or fine-tuning, enabling fine-grained knowledge control in speech AI systems.

Letian Cheng, Junyan Wang, Yan Gao et al.

Perplexity is a widely adopted metric for assessing the predictive quality of large language models (LLMs) and often serves as a reference metric for downstream evaluations. However, recent evidence shows that perplexity can be unreliable, especially when irrelevant long inputs are used, raising concerns for both benchmarking and system deployment. While prior efforts have employed selective input filtering and curated datasets, the impact of input length on perplexity has not been systematically studied from a systems perspective and input length has rarely been treated as a first-class system variable affecting both fairness and efficiency. In this work, we close this gap by introducing LengthBenchmark, a system-conscious evaluation framework that explicitly integrates input length, evaluation protocol design, and system-level costs, evaluating representative LLMs under two scoring protocols (direct accumulation and fixed window sliding) across varying context lengths. Unlike prior work that focuses solely on accuracy-oriented metrics, LengthBenchmark additionally measures latency, memory footprint, and evaluation cost, thereby linking predictive metrics to deployment realities. We further incorporate quantized variants not as a main contribution, but as robustness checks, showing that length-induced biases persist across both full-precision and compressed models. This design disentangles the effects of evaluation logic, quantization, and input length, and demonstrates that length bias is a general phenomenon that undermines fair cross-model comparison. Our analysis yields two key observations: (i) sliding window evaluation consistently inflates performance on short inputs, and (ii) both full-precision and quantized models appear to realise gains as the evaluated segment length grows.

Mujie Liu, Chenze Wang, Liping Chen et al.

The limited availability of labeled brain network data makes it challenging to achieve accurate and interpretable psychiatric diagnoses. While self-supervised learning (SSL) offers a promising solution, existing methods often rely on augmentation strategies that can disrupt crucial structural semantics in brain graphs. To address this, we propose SAM-BG, a two-stage framework for learning brain graph representations with structural semantic preservation. In the pre-training stage, an edge masker is trained on a small labeled subset to capture key structural semantics. In the SSL stage, the extracted structural priors guide a structure-aware augmentation process, enabling the model to learn more semantically meaningful and robust representations. Experiments on two real-world psychiatric datasets demonstrate that SAM-BG outperforms state-of-the-art methods, particularly in small-labeled data settings, and uncovers clinically relevant connectivity patterns that enhance interpretability. Our code is available at https://github.com/mjliu99/SAM-BG.

Siyi Wang, Shihong Tan, Siyi Liu et al.

Emotional expression in human speech is nuanced and compositional, often involving multiple, sometimes conflicting, affective cues that may diverge from linguistic content. In contrast, most expressive text-to-speech systems enforce a single utterance-level emotion, collapsing affective diversity and suppressing mixed or text-emotion-misaligned expression. While activation steering via latent direction vectors offers a promising solution, it remains unclear whether emotion representations are linearly steerable in TTS, where steering should be applied within hybrid TTS architectures, and how such complex emotion behaviors should be evaluated. This paper presents the first systematic analysis of activation steering for emotional control in hybrid TTS models, introducing a quantitative, controllable steering framework, and multi-rater evaluation protocols that enable composable mixed-emotion synthesis and reliable text-emotion mismatch synthesis. Our results demonstrate, for the first time, that emotional prosody and expressive variability are primarily synthesized by the TTS language module instead of the flow-matching module, and also provide a lightweight steering approach for generating natural, human-like emotional speech.

Wenda Zhang, Hongyu Jin, Siyi Wang et al.

Speech Emotion Recognition models typically use single categorical labels, overlooking the inherent ambiguity of human emotions. Ambiguous Emotion Recognition addresses this by representing emotions as probability distributions, but progress is limited by unreliable ground-truth distributions inferred from sparse human annotations. This paper explores whether Large Audio-Language Models (ALMs) can mitigate the annotation bottleneck by generating high-quality synthetic annotations. We introduce a framework leveraging ALMs to create Synthetic Perceptual Proxies, augmenting human annotations to improve ground-truth distribution reliability. We validate these proxies through statistical analysis of their alignment with human distributions and evaluate their impact by fine-tuning ALMs with the augmented emotion distributions. Furthermore, to address class imbalance and enable unbiased evaluation, we propose DiME-Aug, a Distribution-aware Multimodal Emotion Augmentation strategy. Experiments on IEMOCAP and MSP-Podcast show that synthetic annotations enhance emotion distribution, especially in low-ambiguity regions where annotation agreement is high. However, benefits diminish for highly ambiguous emotions with greater human disagreement. This work provides the first evidence that ALMs could address annotation scarcity in ambiguous emotion recognition, but highlights the need for more advanced prompting or generation strategies to handle highly ambiguous cases.

Yu Wu, Ting Dang, Dimitris Spathis et al. · cambridge

Contrastive learning (CL) has emerged as a promising approach for representation learning in time series data by embedding similar pairs closely while distancing dissimilar ones. However, existing CL methods often introduce false negative pairs (FNPs) by neglecting inherent characteristics and then randomly selecting distinct segments as dissimilar pairs, leading to erroneous representation learning, reduced model performance, and overall inefficiency. To address these issues, we systematically define and categorize FNPs in time series into semantic false negative pairs and temporal false negative pairs for the first time: the former arising from overlooking similarities in label categories, which correlates with similarities in non-stationarity and the latter from neglecting temporal proximity. Moreover, we introduce StatioCL, a novel CL framework that captures non-stationarity and temporal dependency to mitigate both FNPs and rectify the inaccuracies in learned representations. By interpreting and differentiating non-stationary states, which reflect the correlation between trends or temporal dynamics with underlying data patterns, StatioCL effectively captures the semantic characteristics and eliminates semantic FNPs. Simultaneously, StatioCL establishes fine-grained similarity levels based on temporal dependencies to capture varying temporal proximity between segments and to mitigate temporal FNPs. Evaluated on real-world benchmark time series classification datasets, StatioCL demonstrates a substantial improvement over state-of-the-art CL methods, achieving a 2.9% increase in Recall and a 19.2% reduction in FNPs. Most importantly, StatioCL also shows enhanced data efficiency and robustness against label scarcity.

Hong Jia, Shiya Fu, Feng Xia et al.

The emergence of Small Language Models (SLMs) as privacy-preserving alternatives for sensitive applications raises a fundamental question about their inherent understanding capabilities compared to Large Language Models (LLMs). This paper investigates the mental health understanding capabilities of current SLMs through systematic evaluation across diverse classification tasks. Employing zero-shot and few-shot learning paradigms, we benchmark their performance against established LLM baselines to elucidate their relative strengths and limitations in this critical domain. We assess five state-of-the-art SLMs (Phi-3, Phi-3.5, Qwen2.5, Llama-3.2, Gemma2) against three LLMs (GPT-4, FLAN-T5-XXL, Alpaca-7B) on six mental health understanding tasks. Our findings reveal that SLMs achieve mean performance within 2\% of LLMs on binary classification tasks (F1 scores of 0.64 vs 0.66 in zero-shot settings), demonstrating notable competence despite orders of magnitude fewer parameters. Both model categories experience similar degradation on multi-class severity tasks (a drop of over 30\%), suggesting that nuanced clinical understanding challenges transcend model scale. Few-shot prompting provides substantial improvements for SLMs (up to 14.6\%), while LLM gains are more variable. Our work highlights the potential of SLMs in mental health understanding, showing they can be effective privacy-preserving tools for analyzing sensitive online text data. In particular, their ability to quickly adapt and specialize with minimal data through few-shot learning positions them as promising candidates for scalable mental health screening tools.

Jiaheng Dong, Hong Jia, Soumyajit Chatterjee et al.

Speech Foundation Models encounter significant performance degradation when deployed in real-world scenarios involving acoustic domain shifts, such as background noise and speaker accents. Test-time adaptation (TTA) has recently emerged as a viable strategy to address such domain shifts at inference time without requiring access to source data or labels. However, existing TTA approaches, particularly those relying on backpropagation, are memory-intensive, limiting their applicability in speech tasks and resource-constrained settings. Although backpropagation-free methods offer improved efficiency, existing ones exhibit poor accuracy. This is because they are predominantly developed for vision tasks, which fundamentally differ from speech task formulations, noise characteristics, and model architecture, posing unique transferability challenges. In this paper, we introduce E-BATS, the first Efficient BAckpropagation-free TTA framework designed explicitly for speech foundation models. E-BATS achieves a balance between adaptation effectiveness and memory efficiency through three key components: (i) lightweight prompt adaptation for a forward-pass-based feature alignment, (ii) a multi-scale loss to capture both global (utterance-level) and local distribution shifts (token-level) and (iii) a test-time exponential moving average mechanism for stable adaptation across utterances. Experiments conducted on four noisy speech datasets spanning sixteen acoustic conditions demonstrate consistent improvements, with 4.1%-13.5% accuracy gains over backpropagation-free baselines and 2.0-6.4 times GPU memory savings compared to backpropagation-based methods. By enabling scalable and robust adaptation under acoustic variability, this work paves the way for developing more efficient adaptation approaches for practical speech processing systems in real-world environments.

Ting Dang, Yan Gao, Hong Jia

Large language models (LLMs) have shown exceptional versatility in natural language processing, prompting recent efforts to extend their multimodal capabilities to speech processing through the development of audio large language models (Audio LLMs). While Audio LLMs excel in tasks such as speech recognition and synthesis, it remains unclear how they perform when faced with the auditory cognitive challenges posed by real-world environments, such as audio comprehension and listening recall, particularly in the presence of background noise or overlapping speech. Unlike text-based LLMs, which have access to vast amounts of text data for pre-training, retraining Audio LLMs with diverse auditory cognitive scenes is difficult due to the limited datasets that simulate real-world auditory cognitive scenarios and the challenge of acquiring auditory cognitive labels for training. While test-time compute (TTC) methods have been shown to enhance the capabilities of text-based LLMs during inference, a key challenge lies in designing these TTC methods to improve the auditory capabilities of Audio LLMs. This study aims to address these two research gaps by: i) exploring the auditory cognitive capabilities of Audio LLMs, and ii) enhancing their capabilities using TTC approaches. We have investigated five different Audio LLMs for auditory cognition using a \textit{self-collected} database and have proposed five TTC approaches to enhance auditory cognitive capabilities during inference. Our findings reveal that Audio LLMs performance decreases in more challenging auditory cognitive tasks. The proposed TTC approaches significantly enhance cognitive auditory capabilities, advancing the development of more adaptable and resilient Audio LLMs for practical applications such as assistive listening devices, voice-based AI assistants, and communication technologies.

Yang Xiao, Aso Mahmudi, Nick Thieberger et al.

Speech foundation models struggle with low-resource Pacific Indigenous languages because of severe data scarcity. Furthermore, full fine-tuning risks catastrophic forgetting. To address this gap, we present an empirical study adapting models to real-world Pacific datasets. We investigate how data volume and linguistic features affect adaptation success. Specifically, we evaluate strategies including Full Fine-Tuning and Low-Rank Adaptation (LoRA). Additionally, we analyze a continual learning framework for sequentially acquiring multiple languages. We demonstrate that adapting to these distant languages causes severe internal representational drift. Consequently, these models face a strict plasticity and stability dilemma. While LoRA adapts well initially, it suffers from catastrophic forgetting during sequential learning. Ultimately, this study highlights the urgent need for robust adaptation strategies tailored to underrepresented languages.

Ting Dang, Soumyajit Chatterjee, Hong Jia et al.

Test time adaptation (TTA) has emerged as a promising solution to adapt pre-trained models to new, unseen data distributions using unlabeled target domain data. However, most TTA methods are designed for independent data, often overlooking the time series data and rarely addressing forecasting tasks. This paper presents AdaNODEs, an innovative source-free TTA method tailored explicitly for time series forecasting. By leveraging Neural Ordinary Differential Equations (NODEs), we propose a novel adaptation framework that accommodates the unique characteristics of distribution shifts in time series data. Moreover, we innovatively propose a new loss function to tackle TTA for forecasting tasks. AdaNODEs only requires updating limited model parameters, showing effectiveness in capturing temporal dependencies while avoiding significant memory usage. Extensive experiments with one- and high-dimensional data demonstrate that AdaNODEs offer relative improvements of 5.88\% and 28.4\% over the SOTA baselines, especially demonstrating robustness across higher severity distribution shifts.

Yang Xiao, Eun-Jung Holden, Ting Dang

Recent speech foundation models excel at multilingual automatic speech recognition (ASR) for high-resource languages, but adapting them to low-resource languages remains challenging due to data scarcity and efficiency constraints. Full-model fine-tuning is computationally expensive and prone to overfitting, while parameter-efficient methods like LoRA apply adaptation uniformly across layers, overlooking internal representations thus compromising effectiveness and efficiency. We analyze multilingual ASR models and reveal a U-shaped adaptability pattern: early and late layers are language-specific and require more adaptation, while intermediate layers retain shared semantics and need less. Building on this observation, we propose DAMA, a Depth-Aware Model Adaptation framework that allocates adaptation capacity according to each layer's role. DAMA also introduces Singular Value Decomposition (SVD)-based initialization to constrain adaptation and preserve the U-shaped pattern, as well as a frozen middle-layer basis for further efficiency. Evaluated on 18 low-resource languages across two benchmark datasets, DAMA matches or surpasses state-of-the-art accuracy with 80% fewer trainable parameters, achieves a 29% error reduction under extreme data scarcity, and significantly improves memory, training time, and computational efficiency over baselines. These results highlight the benefits of structure-aware adaptation for efficient, scalable multilingual ASR.

Manh Pham Hung, Changshuo Hu, Ting Dang et al.

Device-guided music transfer adapts playback across unseen devices for users who lack them. Existing methods mainly focus on modifying the timbre, rhythm, harmony, or instrumentation to mimic genres or artists, overlooking the diverse hardware properties of the playback device (i.e., speaker). Therefore, we propose DeMT, which processes a speaker's frequency response curve as a line graph using a vision-language model to extract device embeddings. These embeddings then condition a hybrid transformer via feature-wise linear modulation. Fine-tuned on a self-collected dataset, DeMT enables effective speaker-style transfer and robust few-shot adaptation for unseen devices, supporting applications like device-style augmentation and quality enhancement.

Xin Wang, Ting Dang, Xinyu Zhang et al.

Mobile and wearable healthcare monitoring play a vital role in facilitating timely interventions, managing chronic health conditions, and ultimately improving individuals' quality of life. Previous studies on large language models (LLMs) have highlighted their impressive generalization abilities and effectiveness in healthcare prediction tasks. However, most LLM-based healthcare solutions are cloud-based, which raises significant privacy concerns and results in increased memory usage and latency. To address these challenges, there is growing interest in compact models, Small Language Models (SLMs), which are lightweight and designed to run locally and efficiently on mobile and wearable devices. Nevertheless, how well these models perform in healthcare prediction remains largely unexplored. We systematically evaluated SLMs on health prediction tasks using zero-shot, few-shot, and instruction fine-tuning approaches, and deployed the best performing fine-tuned SLMs on mobile devices to evaluate their real-world efficiency and predictive performance in practical healthcare scenarios. Our results show that SLMs can achieve performance comparable to LLMs while offering substantial gains in efficiency and privacy. However, challenges remain, particularly in handling class imbalance and few-shot scenarios. These findings highlight SLMs, though imperfect in their current form, as a promising solution for next-generation, privacy-preserving healthcare monitoring.

Yiwen Ye, Yicheng Wu, Xiangde Luo et al.

Foundation models have become a promising paradigm for advancing medical image analysis, particularly for segmentation tasks where downstream applications often emerge sequentially. Existing fine-tuning strategies, however, remain limited: parallel fine-tuning isolates tasks and fails to exploit shared knowledge, while multi-task fine-tuning requires simultaneous access to all datasets and struggles with incremental task integration. To address these challenges, we propose MedSeqFT, a sequential fine-tuning framework that progressively adapts pre-trained models to new tasks while refining their representational capacity. MedSeqFT introduces two core components: (1) Maximum Data Similarity (MDS) selection, which identifies downstream samples most representative of the original pre-training distribution to preserve general knowledge, and (2) Knowledge and Generalization Retention Fine-Tuning (K&G RFT), a LoRA-based knowledge distillation scheme that balances task-specific adaptation with the retention of pre-trained knowledge. Extensive experiments on two multi-task datasets covering ten 3D segmentation tasks demonstrate that MedSeqFT consistently outperforms state-of-the-art fine-tuning strategies, yielding substantial performance gains (e.g., an average Dice improvement of 3.0%). Furthermore, evaluations on two unseen tasks (COVID-19-20 and Kidney) verify that MedSeqFT enhances transferability, particularly for tumor segmentation. Visual analyses of loss landscapes and parameter variations further highlight the robustness of MedSeqFT. These results establish sequential fine-tuning as an effective, knowledge-retentive paradigm for adapting foundation models to evolving clinical tasks. Code will be released.

Mujie Liu, Mengchu Zhu, Qichao Dong et al.

Psychiatric disorders involve complex neural activity changes, with functional magnetic resonance imaging (fMRI) data serving as key diagnostic evidence. However, data scarcity and the diverse nature of fMRI information pose significant challenges. While graph-based self-supervised learning (SSL) methods have shown promise in brain network analysis, they primarily focus on time-domain representations, often overlooking the rich information embedded in the frequency domain. To overcome these limitations, we propose Frequency-Enhanced Network (FENet), a novel SSL framework specially designed for fMRI data that integrates time-domain and frequency-domain information to improve psychiatric disorder detection in small-sample datasets. FENet constructs multi-view brain networks based on the inherent properties of fMRI data, explicitly incorporating frequency information into the learning process of representation. Additionally, it employs domain-specific encoders to capture temporal-spectral characteristics, including an efficient frequency-domain encoder that highlights disease-relevant frequency features. Finally, FENet introduces a domain consistency-guided learning objective, which balances the utilization of diverse information and generates frequency-enhanced brain graph representations. Experiments on two real-world medical datasets demonstrate that FENet outperforms state-of-the-art methods while maintaining strong performance in minimal data conditions. Furthermore, we analyze the correlation between various frequency-domain features and psychiatric disorders, emphasizing the critical role of high-frequency information in disorder detection.

Ting Dang, Jing Han, Tong Xia et al.

Recent work has shown the potential of using audio data (eg, cough, breathing, and voice) in the screening for COVID-19. However, these approaches only focus on one-off detection and detect the infection given the current audio sample, but do not monitor disease progression in COVID-19. Limited exploration has been put forward to continuously monitor COVID-19 progression, especially recovery, through longitudinal audio data. Tracking disease progression characteristics could lead to more timely treatment. The primary objective of this study is to explore the potential of longitudinal audio samples over time for COVID-19 progression prediction and, especially, recovery trend prediction using sequential deep learning techniques. Crowdsourced respiratory audio data, including breathing, cough, and voice samples, from 212 individuals over 5-385 days were analyzed. We developed a deep learning-enabled tracking tool using gated recurrent units (GRUs) to detect COVID-19 progression by exploring the audio dynamics of the individuals' historical audio biomarkers. The investigation comprised 2 parts: (1) COVID-19 detection in terms of positive and negative (healthy) tests, and (2) longitudinal disease progression prediction over time in terms of probability of positive tests. The strong performance for COVID-19 detection, yielding an AUROC of 0.79, a sensitivity of 0.75, and a specificity of 0.71 supported the effectiveness of the approach compared to methods that do not leverage longitudinal dynamics. We further examined the predicted disease progression trajectory, displaying high consistency with test results with a correlation of 0.75 in the test cohort and 0.86 in a subset of the test cohort who reported recovery. Our findings suggest that monitoring COVID-19 evolution via longitudinal audio data has potential in the tracking of individuals' disease progression and recovery.

Kayla-Jade Butkow, Ting Dang, Andrea Ferlini et al.

With the soaring adoption of in-ear wearables, the research community has started investigating suitable in-ear heart rate (HR) detection systems. HR is a key physiological marker of cardiovascular health and physical fitness. Continuous and reliable HR monitoring with wearable devices has therefore gained increasing attention in recent years. Existing HR detection systems in wearables mainly rely on photoplethysmography (PPG) sensors, however, these are notorious for poor performance in the presence of human motion. In this work, leveraging the occlusion effect that enhances low-frequency bone-conducted sounds in the ear canal, we investigate for the first time \textit{in-ear audio-based motion-resilient} HR monitoring. We first collected HR-induced sounds in the ear canal leveraging an in-ear microphone under stationary and three different activities (i.e., walking, running, and speaking). Then, we devised a novel deep learning based motion artefact (MA) mitigation framework to denoise the in-ear audio signals, followed by an HR estimation algorithm to extract HR. With data collected from 20 subjects over four activities, we demonstrate that hEARt, our end-to-end approach, achieves a mean absolute error (MAE) of 3.02 $\pm$ 2.97~BPM, 8.12 $\pm$ 6.74~BPM, 11.23 $\pm$ 9.20~BPM and 9.39 $\pm$ 6.97~BPM for stationary, walking, running and speaking, respectively, opening the door to a new non-invasive and affordable HR monitoring with usable performance for daily activities. Not only does hEARt outperform previous in-ear HR monitoring work, but it outperforms reported in-ear PPG performance.

Jingyao Wu, Ting Dang, Vidhyasaharan Sethu et al.

There is growing interest in affective computing for the representation and prediction of emotions along ordinal scales. However, the term ordinal emotion label has been used to refer to both absolute notions such as low or high arousal, as well as relation notions such as arousal is higher at one instance compared to another. In this paper, we introduce the terminology absolute and relative ordinal labels to make this distinction clear and investigate both with a view to integrate them and exploit their complementary nature. We propose a Markovian framework referred to as Dynamic Ordinal Markov Model (DOMM) that makes use of both absolute and relative ordinal information, to improve speech based ordinal emotion prediction. Finally, the proposed framework is validated on two speech corpora commonly used in affective computing, the RECOLA and the IEMOCAP databases, across a range of system configurations. The results consistently indicate that integrating relative ordinal information improves absolute ordinal emotion prediction.

Jing Han, Tong Xia, Dimitris Spathis et al.

Researchers have been battling with the question of how we can identify Coronavirus disease (COVID-19) cases efficiently, affordably and at scale. Recent work has shown how audio based approaches, which collect respiratory audio data (cough, breathing and voice) can be used for testing, however there is a lack of exploration of how biases and methodological decisions impact these tools' performance in practice. In this paper, we explore the realistic performance of audio-based digital testing of COVID-19. To investigate this, we collected a large crowdsourced respiratory audio dataset through a mobile app, alongside recent COVID-19 test result and symptoms intended as a ground truth. Within the collected dataset, we selected 5,240 samples from 2,478 participants and split them into different participant-independent sets for model development and validation. Among these, we controlled for potential confounding factors (such as demographics and language). The unbiased model takes features extracted from breathing, coughs, and voice signals as predictors and yields an AUC-ROC of 0.71 (95\% CI: 0.65$-$0.77). We further explore different unbalanced distributions to show how biases and participant splits affect performance. Finally, we discuss how the realistic model presented could be integrated in clinical practice to realize continuous, ubiquitous, sustainable and affordable testing at population scale.

Tong Xia, Jing Han, Lorena Qendro et al.

Recently, sound-based COVID-19 detection studies have shown great promise to achieve scalable and prompt digital pre-screening. However, there are still two unsolved issues hindering the practice. First, collected datasets for model training are often imbalanced, with a considerably smaller proportion of users tested positive, making it harder to learn representative and robust features. Second, deep learning models are generally overconfident in their predictions. Clinically, false predictions aggravate healthcare costs. Estimation of the uncertainty of screening would aid this. To handle these issues, we propose an ensemble framework where multiple deep learning models for sound-based COVID-19 detection are developed from different but balanced subsets from original data. As such, data are utilized more effectively compared to traditional up-sampling and down-sampling approaches: an AUC of 0.74 with a sensitivity of 0.68 and a specificity of 0.69 is achieved. Simultaneously, we estimate uncertainty from the disagreement across multiple models. It is shown that false predictions often yield higher uncertainty, enabling us to suggest the users with certainty higher than a threshold to repeat the audio test on their phones or to take clinical tests if digital diagnosis still fails. This study paves the way for a more robust sound-based COVID-19 automated screening system.