Shrikanth Narayanan

Tiantian Feng, Digbalay Bose, Tuo Zhang et al. · amazon-science

Over the past few years, Federated Learning (FL) has become an emerging machine learning technique to tackle data privacy challenges through collaborative training. In the Federated Learning algorithm, the clients submit a locally trained model, and the server aggregates these parameters until convergence. Despite significant efforts that have been made to FL in fields like computer vision, audio, and natural language processing, the FL applications utilizing multimodal data streams remain largely unexplored. It is known that multimodal learning has broad real-world applications in emotion recognition, healthcare, multimedia, and social media, while user privacy persists as a critical concern. Specifically, there are no existing FL benchmarks targeting multimodal applications or related tasks. In order to facilitate the research in multimodal FL, we introduce FedMultimodal, the first FL benchmark for multimodal learning covering five representative multimodal applications from ten commonly used datasets with a total of eight unique modalities. FedMultimodal offers a systematic FL pipeline, enabling end-to-end modeling framework ranging from data partition and feature extraction to FL benchmark algorithms and model evaluation. Unlike existing FL benchmarks, FedMultimodal provides a standardized approach to assess the robustness of FL against three common data corruptions in real-life multimodal applications: missing modalities, missing labels, and erroneous labels. We hope that FedMultimodal can accelerate numerous future research directions, including designing multimodal FL algorithms toward extreme data heterogeneity, robustness multimodal FL, and efficient multimodal FL. The datasets and benchmark results can be accessed at: https://github.com/usc-sail/fed-multimodal.

Tiantian Feng, Anfeng Xu, Xuan Shi et al.

We present ChildVox, a novel benchmark for characterizing the diverse acoustic signals through which children communicate. Specifically, ChildVox follows the full developmental trajectory from birth through school age, covering physiological sounds, non-linguistic vocalizations, canonical syllables, and spoken language. ChildVox integrates more than 20 sub-tasks across 17 child-centered audio and speech datasets, enabling systematic cross-corpus and cross-domain comparison. We evaluate a representative range of audio and speech foundation models, including self-supervised, ASR-oriented, and large audio-language models, on tasks including physiological sound classification, vocalization and canonical syllables modeling, and speech quality assessment and recognition. Benchmark results show that ChildVox provides a suite of high-performance models in recognizing a wide range of acoustic signals from children, supporting downstream applications such as characterizing children's language levels and tracking speech production with age.

Georgios Chochlakis, Tejas Srinivasan, Jesse Thomason et al. · uw

We propose the Vision-and-Augmented-Language Transformer (VAuLT). VAuLT is an extension of the popular Vision-and-Language Transformer (ViLT), and improves performance on vision-and-language (VL) tasks that involve more complex text inputs than image captions while having minimal impact on training and inference efficiency. ViLT, importantly, enables efficient training and inference in VL tasks, achieved by encoding images using a linear projection of patches instead of an object detector. However, it is pretrained on captioning datasets, where the language input is simple, literal, and descriptive, therefore lacking linguistic diversity. So, when working with multimedia data in the wild, such as multimodal social media data, there is a notable shift from captioning language data, as well as diversity of tasks. We indeed find evidence that the language capacity of ViLT is lacking. The key insight and novelty of VAuLT is to propagate the output representations of a large language model (LM) like BERT to the language input of ViLT. We show that joint training of the LM and ViLT can yield relative improvements up to 20% over ViLT and achieve state-of-the-art or comparable performance on VL tasks involving richer language inputs and affective constructs, such as for Target-Oriented Sentiment Classification in TWITTER-2015 and TWITTER-2017, and Sentiment Classification in MVSA-Single and MVSA-Multiple. Our code is available at https://github.com/gchochla/VAuLT.

Kleanthis Avramidis, Dominika Kunc, Bartosz Perz et al.

Ubiquitous sensing from wearable devices in the wild holds promise for enhancing human well-being, from diagnosing clinical conditions and measuring stress to building adaptive health promoting scaffolds. But the large volumes of data therein across heterogeneous contexts pose challenges for conventional supervised learning approaches. Representation Learning from biological signals is an emerging realm catalyzed by the recent advances in computational modeling and the abundance of publicly shared databases. The electrocardiogram (ECG) is the primary researched modality in this context, with applications in health monitoring, stress and affect estimation. Yet, most studies are limited by small-scale controlled data collection and over-parameterized architecture choices. We introduce \textbf{WildECG}, a pre-trained state-space model for representation learning from ECG signals. We train this model in a self-supervised manner with 275,000 10s ECG recordings collected in the wild and evaluate it on a range of downstream tasks. The proposed model is a robust backbone for ECG analysis, providing competitive performance on most of the tasks considered, while demonstrating efficacy in low-resource regimes. The code and pre-trained weights are shared publicly at https://github.com/klean2050/tiles_ecg_model.

Nikolaos Antoniou, Athanasios Katsamanis, Theodoros Giannakopoulos et al.

There is an imminent need for guidelines and standard test sets to allow direct and fair comparisons of speech emotion recognition (SER). While resources, such as the Interactive Emotional Dyadic Motion Capture (IEMOCAP) database, have emerged as widely-adopted reference corpora for researchers to develop and test models for SER, published work reveals a wide range of assumptions and variety in its use that challenge reproducibility and generalization. Based on a critical review of the latest advances in SER using IEMOCAP as the use case, our work aims at two contributions: First, using an analysis of the recent literature, including assumptions made and metrics used therein, we provide a set of SER evaluation guidelines. Second, using recent publications with open-sourced implementations, we focus on reproducibility assessment in SER.

Georgios Chochlakis, Gireesh Mahajan, Sabyasachee Baruah et al.

Detecting emotions expressed in text has become critical to a range of fields. In this work, we investigate ways to exploit label correlations in multi-label emotion recognition models to improve emotion detection. First, we develop two modeling approaches to the problem in order to capture word associations of the emotion words themselves, by either including the emotions in the input, or by leveraging Masked Language Modeling (MLM). Second, we integrate pairwise constraints of emotion representations as regularization terms alongside the classification loss of the models. We split these terms into two categories, local and global. The former dynamically change based on the gold labels, while the latter remain static during training. We demonstrate state-of-the-art performance across Spanish, English, and Arabic in SemEval 2018 Task 1 E-c using monolingual BERT-based models. On top of better performance, we also demonstrate improved robustness. Code is available at https://github.com/gchochla/Demux-MEmo.

Junhyeok Lee, Xiluo He, Jihwan Lee et al.

Neural audio codecs optimized for mel-spectrogram reconstruction often fail to preserve intelligibility. While semantic encoder distillation improves encoded representations, it does not guarantee content preservation in reconstructed speech. In this work, we demonstrate that self-supervised representation reconstruction (SSRR) loss fundamentally improves codec training and performance. First, SSRR significantly accelerates convergence, enabling competitive results using only a single GPU. Second, it enhances intelligibility by reconstructing distilled self-supervised representations from codec outputs. Third, SSRR enables high intelligibility without additional lookahead in streaming Transformer-based codecs, allowing a zero-lookahead architecture for real-time deployment. As a result, our JHCodec achieves state-of-the-art performance while maintaining minimal latency and reduced training cost. We open-source the full implementation, training pipeline, and demo on Github https://github.com/jhcodec843/jhcodec.

Rajat Hebbar, Digbalay Bose, Krishna Somandepalli et al.

Audio event detection is a widely studied audio processing task, with applications ranging from self-driving cars to healthcare. In-the-wild datasets such as Audioset have propelled research in this field. However, many efforts typically involve manual annotation and verification, which is expensive to perform at scale. Movies depict various real-life and fictional scenarios which makes them a rich resource for mining a wide-range of audio events. In this work, we present a dataset of audio events called Subtitle-Aligned Movie Sounds (SAM-S). We use publicly-available closed-caption transcripts to automatically mine over 110K audio events from 430 movies. We identify three dimensions to categorize audio events: sound, source, quality, and present the steps involved to produce a final taxonomy of 245 sounds. We discuss the choices involved in generating the taxonomy, and also highlight the human-centered nature of sounds in our dataset. We establish a baseline performance for audio-only sound classification of 34.76% mean average precision and show that incorporating visual information can further improve the performance by about 5%. Data and code are made available for research at https://github.com/usc-sail/mica-subtitle-aligned-movie-sounds

Georgios Chochlakis, Gireesh Mahajan, Sabyasachee Baruah et al.

The need for emotional inference from text continues to diversify as more and more disciplines integrate emotions into their theories and applications. These needs include inferring different emotion types, handling multiple languages, and different annotation formats. A shared model between different configurations would enable the sharing of knowledge and a decrease in training costs, and would simplify the process of deploying emotion recognition models in novel environments. In this work, we study how we can build a single model that can transition between these different configurations by leveraging multilingual models and Demux, a transformer-based model whose input includes the emotions of interest, enabling us to dynamically change the emotions predicted by the model. Demux also produces emotion embeddings, and performing operations on them allows us to transition to clusters of emotions by pooling the embeddings of each cluster. We show that Demux can simultaneously transfer knowledge in a zero-shot manner to a new language, to a novel annotation format and to unseen emotions. Code is available at https://github.com/gchochla/Demux-MEmo .

Rahul Sharma, Shrikanth Narayanan

Active speaker detection in videos addresses associating a source face, visible in the video frames, with the underlying speech in the audio modality. The two primary sources of information to derive such a speech-face relationship are i) visual activity and its interaction with the speech signal and ii) co-occurrences of speakers' identities across modalities in the form of face and speech. The two approaches have their limitations: the audio-visual activity models get confused with other frequently occurring vocal activities, such as laughing and chewing, while the speakers' identity-based methods are limited to videos having enough disambiguating information to establish a speech-face association. Since the two approaches are independent, we investigate their complementary nature in this work. We propose a novel unsupervised framework to guide the speakers' cross-modal identity association with the audio-visual activity for active speaker detection. Through experiments on entertainment media videos from two benchmark datasets, the AVA active speaker (movies) and Visual Person Clustering Dataset (TV shows), we show that a simple late fusion of the two approaches enhances the active speaker detection performance.

Kleanthis Avramidis, Mohammad Rostami, Melinda Chang et al.

Papilledema is an ophthalmic neurologic disorder in which increased intracranial pressure leads to swelling of the optic nerves. Undiagnosed papilledema in children may lead to blindness and may be a sign of life-threatening conditions, such as brain tumors. Robust and accurate clinical diagnosis of this syndrome can be facilitated by automated analysis of fundus images using deep learning, especially in the presence of challenges posed by pseudopapilledema that has similar fundus appearance but distinct clinical implications. We present a deep learning-based algorithm for the automatic detection of pediatric papilledema. Our approach is based on optic disc localization and detection of explainable papilledema indicators through data augmentation. Experiments on real-world clinical data demonstrate that our proposed method is effective with a diagnostic accuracy comparable to expert ophthalmologists.

Rahul Sharma, Shrikanth Narayanan

An essential goal of computational media intelligence is to support understanding how media stories -- be it news, commercial or entertainment media -- represent and reflect society and these portrayals are perceived. People are a central element of media stories. This paper focuses on understanding the representation and depiction of background characters in media depictions, primarily movies and TV shows. We define the background characters as those who do not participate vocally in any scene throughout the movie and address the problem of localizing background characters in videos. We use an active speaker localization system to extract high-confidence face-speech associations and generate audio-visual profiles for talking characters in a movie by automatically clustering them. Using a face verification system, we then prune all the face-tracks which match any of the generated character profiles and obtain the background character face-tracks. We curate a background character dataset which provides annotations for background character for a set of TV shows, and use it to evaluate the performance of the background character detection framework.

Zhuohao Chen, Nikolaos Flemotomos, Zac E. Imel et al.

In psychotherapy interactions, the quality of a session is assessed by codifying the communicative behaviors of participants during the conversation through manual observation and annotation. Developing computational approaches for automated behavioral coding can reduce the burden on human coders and facilitate the objective evaluation of the intervention. In the real world, however, implementing such algorithms is associated with data sparsity challenges since privacy concerns lead to limited available in-domain data. In this paper, we leverage a publicly available conversation-based dataset and transfer knowledge to the low-resource behavioral coding task by performing an intermediate language model training via meta-learning. We introduce a task augmentation method to produce a large number of "analogy tasks" - tasks similar to the target one - and demonstrate that the proposed framework predicts target behaviors more accurately than all the other baseline models.

Kleanthis Avramidis, Kranti Adsul, Digbalay Bose et al.

This paper presents the approach and results of USC SAIL's submission to the Signal Processing Grand Challenge 2023 - e-Prevention (Task 2), on detecting relapses in psychotic patients. Relapse prediction has proven to be challenging, primarily due to the heterogeneity of symptoms and responses to treatment between individuals. We address these challenges by investigating the use of sleep behavior features to estimate relapse days as outliers in an unsupervised machine learning setting. We extract informative features from human activity and heart rate data collected in the wild, and evaluate various combinations of feature types and time resolutions. We found that short-time sleep behavior features outperformed their awake counterparts and larger time intervals. Our submission was ranked 3rd in the Task's official leaderboard, demonstrating the potential of such features as an objective and non-invasive predictor of psychotic relapses.

Rahul Sharma, Shrikanth Narayanan

We present a cross-modal unsupervised framework for active speaker detection in media content such as TV shows and movies. Machine learning advances have enabled impressive performance in identifying individuals from speech and facial images. We leverage speaker identity information from speech and faces, and formulate active speaker detection as a speech-face assignment task such that the active speaker's face and the underlying speech identify the same person (character). We express the speech segments in terms of their associated speaker identity distances, from all other speech segments, to capture a relative identity structure for the video. Then we assign an active speaker's face to each speech segment from the concurrently appearing faces such that the obtained set of active speaker faces displays a similar relative identity structure. Furthermore, we propose a simple and effective approach to address speech segments where speakers are present off-screen. We evaluate the proposed system on three benchmark datasets -- Visual Person Clustering dataset, AVA-active speaker dataset, and Columbia dataset -- consisting of videos from entertainment and broadcast media, and show competitive performance to state-of-the-art fully supervised methods.

Victor Ardulov, Torrey A. Creed, David C. Atkins et al.

In order to increase mental health equity among the most vulnerable and marginalized communities, it is important to increase access to high-quality therapists. One facet of addressing these needs, is to provide timely feedback to clinicians as they interact with their clients, in a way that is also contextualized to specific clients and interactions they have had. Dynamical systems provide a framework through which to analyze interactions. The present work applies these methods to the domain of automated psychotherapist evaluation for Cognitive Behavioral Therapy (CBT). Our methods extract local dynamic modes from short windows of conversation and learns to correlate the observed dynamics to CBT competence. The results demonstrate the value of this paradigm and outlines the way in which these methods can be used to study and improve therapeutic strategies.

Georgios Chochlakis, Niyantha Maruthu Pandiyan, Kristina Lerman et al.

In-Context Learning (ICL) in Large Language Models (LLM) has emerged as the dominant technique for performing natural language tasks, as it does not require updating the model parameters with gradient-based methods. ICL promises to "adapt" the LLM to perform the present task at a competitive or state-of-the-art level at a fraction of the computational cost. ICL can be augmented by incorporating the reasoning process to arrive at the final label explicitly in the prompt, a technique called Chain-of-Thought (CoT) prompting. However, recent work has found that ICL relies mostly on the retrieval of task priors and less so on "learning" to perform tasks, especially for complex subjective domains like emotion and morality, where priors ossify posterior predictions. In this work, we examine whether "enabling" reasoning also creates the same behavior in LLMs, wherein the format of CoT retrieves reasoning priors that remain relatively unchanged despite the evidence in the prompt. We find that, surprisingly, CoT indeed suffers from the same posterior collapse as ICL for larger language models. Code is avalaible at https://github.com/gchochla/cot-priors.

Yoonsoo Nam, Adam Lehavi, Daniel Yang et al. · allen-ai

Video summarization remains a huge challenge in computer vision due to the size of the input videos to be summarized. We propose an efficient, language-only video summarizer that achieves competitive accuracy with high data efficiency. Using only textual captions obtained via a zero-shot approach, we train a language transformer model and forego image representations. This method allows us to perform filtration amongst the representative text vectors and condense the sequence. With our approach, we gain explainability with natural language that comes easily for human interpretation and textual summaries of the videos. An ablation study that focuses on modality and data compression shows that leveraging text modality only effectively reduces input data processing while retaining comparable results.

Digbalay Bose, Rajat Hebbar, Krishna Somandepalli et al.

Longform media such as movies have complex narrative structures, with events spanning a rich variety of ambient visual scenes. Domain specific challenges associated with visual scenes in movies include transitions, person coverage, and a wide array of real-life and fictional scenarios. Existing visual scene datasets in movies have limited taxonomies and don't consider the visual scene transition within movie clips. In this work, we address the problem of visual scene recognition in movies by first automatically curating a new and extensive movie-centric taxonomy of 179 scene labels derived from movie scripts and auxiliary web-based video datasets. Instead of manual annotations which can be expensive, we use CLIP to weakly label 1.12 million shots from 32K movie clips based on our proposed taxonomy. We provide baseline visual models trained on the weakly labeled dataset called MovieCLIP and evaluate them on an independent dataset verified by human raters. We show that leveraging features from models pretrained on MovieCLIP benefits downstream tasks such as multi-label scene and genre classification of web videos and movie trailers.

Tiantian Feng, Brandon M Booth, Shrikanth Narayanan

Continuously-worn wearable sensors enable researchers to collect copious amounts of rich bio-behavioral time series recordings of real-life activities of daily living, offering unprecedented opportunities to infer novel human behavior patterns during daily routines. Existing approaches to routine discovery through bio-behavioral data rely either on pre-defined notions of activities or use additional non-behavioral measurements as contexts, such as GPS location or localization within the home, presenting risks to user privacy. In this work, we propose a novel wearable time-series mining framework, Hawkes point process On Time series clusters for ROutine Discovery (HOT-ROD), for uncovering behavioral routines from completely unlabeled wearable recordings. We utilize a covariance-based method to generate time-series clusters and discover routines via the Hawkes point process learning algorithm. We empirically validate our approach for extracting routine behaviors using a completely unlabeled time-series collected continuously from over 100 individuals both in and outside of the workplace during a period of ten weeks. Furthermore, we demonstrate this approach intuitively captures daily transitional relationships between physical activity states without using prior knowledge. We also show that the learned behavioral patterns can assist in illuminating an individual's personality and affect.

Aditya Kommineni, Emily Zhou, Kleanthis Avramidis et al.

Evaluating foundation models under appropriate adaptation settings is essential for understanding the quality and transferability of the learned representations. Recent EEG foundation models have demonstrated promising transfer capabilities across tasks and datasets, motivating their growing use in neurotechnology and clinical applications. However, these models are typically evaluated under full fine-tuning on well-curated downstream datasets, a setting that does not reflect biomedical domain constraints such as limited labeled data, reduced sensor coverage, or parameter-efficient adaptation. In this work, we propose a multi-dimensional evaluation framework for assessing EEG models under realistic low-resource conditions. Empirical analysis of both supervised EEG models and recent EEG foundation models, including LaBraM, CSBrain, and CBraMod, across 6 different datasets is performed under the proposed multi-dimensional evaluation framework. We find that EEG foundation models consistently provide performance gains on long-context tasks such as sleep stage prediction and mental health state classification. In contrast, for short-window Brain Computer Interface style tasks, supervised models achieve comparable despite having substantially fewer parameters. Additional analyses demonstrate that current foundation models provide limited robustness to short-window tasks and channel constrained settings. Together, these findings motivate the use of multi-dimensional evaluation protocols that characterize model behavior under realistic use constraints.

Tiantian Feng, Shrikanth Narayanan

Speech Emotion Recognition (SER) application is frequently associated with privacy concerns as it often acquires and transmits speech data at the client-side to remote cloud platforms for further processing. These speech data can reveal not only speech content and affective information but the speaker's identity, demographic traits, and health status. Federated learning (FL) is a distributed machine learning algorithm that coordinates clients to train a model collaboratively without sharing local data. This algorithm shows enormous potential for SER applications as sharing raw speech or speech features from a user's device is vulnerable to privacy attacks. However, a major challenge in FL is limited availability of high-quality labeled data samples. In this work, we propose a semi-supervised federated learning framework, Semi-FedSER, that utilizes both labeled and unlabeled data samples to address the challenge of limited labeled data samples in FL. We show that our Semi-FedSER can generate desired SER performance even when the local label rate l=20 using two SER benchmark datasets: IEMOCAP and MSP-Improv.

Woojae Jeong, Wenhui Cui, Kleanthis Avramidis et al.

Electroencephalography (EEG) offers detailed access to neural dynamics but remains constrained by noise and trial-by-trial variability, limiting decoding performance in data-restricted or complex paradigms. Data augmentation is often employed to enhance feature representations, yet conventional uniform averaging overlooks differences in trial informativeness and can degrade representational quality. We introduce a weighted bootstrapping approach that prioritizes more reliable trials to generate higher-quality augmented samples. In a Sentence Evaluation paradigm, weights were computed from relative ERP differences and applied during probabilistic sampling and averaging. Across conditions, weighted bootstrapping improved decoding accuracy relative to unweighted (from 68.35% to 71.25% at best), demonstrating that emphasizing reliable trials strengthens representational quality. The results demonstrate that reliability-based augmentation yields more robust and discriminative EEG representations. The code is publicly available at https://github.com/lyricists/NeuroBootstrap.

Aditya Kommineni, Emily Zhou, Kleanthis Avramidis et al.

EEG foundation models, pre-trained on large-scale unlabelled EEG data, have emerged as a promising direction towards learning generalizable EEG representations. Despite showing positive results in data-rich regimes, they often fail to outperform significantly smaller supervised models in low-resource settings compared to fully supervised models. We provide a mechanistic account of this shortcoming, attributing it to a fundamental mismatch between reconstruction-based pretext tasks and the idiosyncratic spectral structure of EEG signals, which decompose into distinct high-power aperiodic and low-power oscillatory components. Using controlled, synthetically-generated EEG inputs, we demonstrate that EEG foundation model embeddings are biased to capture the aperiodic components of the EEG signal while under-representing oscillatory components, particularly at higher frequencies. Additionally, linear probe evaluations on real-world BCI datasets further reveal that embeddings encode subject identity more strongly than task-relevant information, thereby reinforcing the low-frequency and aperiodic component bias in foundation model embeddings trained primarily on reconstruction based objectives. Together, these findings elucidate a failure mode in reconstruction based EEG foundation models and motivate future work to incorporate auxiliary losses explicitly targeting high-frequency oscillatory structure as a path toward more capable and generalizable EEG representations.

Digbalay Bose, Rajat Hebbar, Tiantian Feng et al.

Advertisement videos (ads) play an integral part in the domain of Internet e-commerce as they amplify the reach of particular products to a broad audience or can serve as a medium to raise awareness about specific issues through concise narrative structures. The narrative structures of advertisements involve several elements like reasoning about the broad content (topic and the underlying message) and examining fine-grained details involving the transition of perceived tone due to the specific sequence of events and interaction among characters. In this work, to facilitate the understanding of advertisements along the three important dimensions of topic categorization, perceived tone transition, and social message detection, we introduce a multimodal multilingual benchmark called MM-AU composed of over 8.4K videos (147 hours) curated from multiple web sources. We explore multiple zero-shot reasoning baselines through the application of large language models on the ads transcripts. Further, we demonstrate that leveraging signals from multiple modalities, including audio, video, and text, in multimodal transformer-based supervised models leads to improved performance compared to unimodal approaches.

Marcus Ma, Emily Zhou, Leonard Ludwig et al.

We study when, where, and why 978 Holocaust survivors smile in video testimonies. We create an automatic smile detection model from facial features with an F1 of 85% and annotate detected smiles under two established taxonomies of smiling. We produce narrative features on 1,083,417 transcript sentences as well as emotional valence from three different modalities: audio, eye gaze, and text transcript. Smiling rates are significantly correlated with specific semantic topics, narrative structures, and temporal syntaxes across the entire corpus. Smiles often occur during periods of intense negative affect; these negative-affect smiles improve the valence trajectory of surrounding sentences significantly across all three modalities. Smiling reduces eye dynamics and blink rates, and the strength of both of these effects is also modulated by narrative valence. Taken together, we conclude that smiling plays a critical role in regulating emotion and social interaction during traumatic recollection.

Aditya Kommineni, Digbalay Bose, Tiantian Feng et al.

Clinical videos in the context of Autism Spectrum Disorder are often long-form interactions between children and caregivers/clinical professionals, encompassing complex verbal and non-verbal behaviors. Objective analyses of these videos could provide clinicians and researchers with nuanced insights into the behavior of children with Autism Spectrum Disorder. Manually coding these videos is a time-consuming task and requires a high level of domain expertise. Hence, the ability to capture these interactions computationally can augment the manual effort and enable supporting the diagnostic procedure. In this work, we investigate the use of foundation models across three modalities: speech, video, and text, to analyse child-focused interaction sessions. We propose a unified methodology to combine multiple modalities by using large language models as reasoning agents. We evaluate their performance on two tasks with different information granularity: activity recognition and abnormal behavior detection. We find that the proposed multimodal pipeline provides robustness to modality-specific limitations and improves performance on the clinical video analysis compared to unimodal settings.

Nicholas Mehlman, Anirudh Sreeram, Raghuveer Peri et al.

A variety of recent works have looked into defenses for deep neural networks against adversarial attacks particularly within the image processing domain. Speech processing applications such as automatic speech recognition (ASR) are increasingly relying on deep learning models, and so are also prone to adversarial attacks. However, many of the defenses explored for ASR simply adapt the image-domain defenses, which may not provide optimal robustness. This paper explores speech specific defenses using the mel spectral domain, and introduces a novel defense method called 'mel domain noise flooding' (MDNF). MDNF applies additive noise to the mel spectrogram of a speech utterance prior to re-synthesising the audio signal. We test the defenses against strong white-box adversarial attacks such as projected gradient descent (PGD) and Carlini-Wagner (CW) attacks, and show better robustness compared to a randomized smoothing baseline across strong threat models.

Hong Nguyen, Sean Foley, Kevin Huang et al.

Understanding speech production both visually and kinematically can inform second language learning system designs, as well as the creation of speaking characters in video games and animations. In this work, we introduce a data-driven method to visually represent articulator motion in Magnetic Resonance Imaging (MRI) videos of the human vocal tract during speech based on arbitrary audio or speech input. We leverage large pre-trained speech models, which are embedded with prior knowledge, to generalize the visual domain to unseen data using a speech-to-video diffusion model. Our findings demonstrate that the visual generation significantly benefits from the pre-trained speech representations. We also observed that evaluating phonemes in isolation is challenging but becomes more straightforward when assessed within the context of spoken words. Limitations of the current results include the presence of unsmooth tongue motion and video distortion when the tongue contacts the palate.

Digbalay Bose, Rajat Hebbar, Krishna Somandepalli et al.

The process of human affect understanding involves the ability to infer person specific emotional states from various sources including images, speech, and language. Affect perception from images has predominantly focused on expressions extracted from salient face crops. However, emotions perceived by humans rely on multiple contextual cues including social settings, foreground interactions, and ambient visual scenes. In this work, we leverage pretrained vision-language (VLN) models to extract descriptions of foreground context from images. Further, we propose a multimodal context fusion (MCF) module to combine foreground cues with the visual scene and person-based contextual information for emotion prediction. We show the effectiveness of our proposed modular design on two datasets associated with natural scenes and TV shows.

Efthymios Tsaprazlis, Tiantian Feng, Anil Ramakrishna et al.

Existing visual privacy benchmarks largely treat privacy as a binary property, labeling images as private or non-private based on visible sensitive content. We argue that privacy is fundamentally compositional. Attributes that are benign in isolation may combine to produce severe privacy violations. We introduce the Compositional Privacy Risk Taxonomy (CPRT), a regulation-aware framework that organizes visual attributes according to standalone identifiability and compositional harm potential. CPRT defines four graded severity levels and is paired with an interpretable scoring function that assigns continuous privacy severity scores. We further construct a taxonomy-aligned dataset of 6.7K images and derive ground-truth compositional risk scores. By evaluating frontier and open-weight VLMs we find that frontier models align well with compositional severity when provided structured guidance, but systematically underestimate composition-driven risks. Smaller models struggle to internalize graded privacy reasoning. To bridge this gap, we introduce a deployable 8B supervised fine-tuned (SFT) model that closely matches frontier-level performance on compositional privacy assessment.

Tiantian Feng, Anfeng Xu, Xuan Shi et al.

Autism spectrum disorder (ASD) is a neurodevelopmental condition characterized by challenges in social communication, repetitive behavior, and sensory processing. One important research area in ASD is evaluating children's behavioral changes over time during treatment. The standard protocol with this objective is BOSCC, which involves dyadic interactions between a child and clinicians performing a pre-defined set of activities. A fundamental aspect of understanding children's behavior in these interactions is automatic speech understanding, particularly identifying who speaks and when. Conventional approaches in this area heavily rely on speech samples recorded from a spectator perspective, and there is limited research on egocentric speech modeling. In this study, we design an experiment to perform speech sampling in BOSCC interviews from an egocentric perspective using wearable sensors and explore pre-training Ego4D speech samples to enhance child-adult speaker classification in dyadic interactions. Our findings highlight the potential of egocentric speech collection and pre-training to improve speaker classification accuracy.

Aditya Ashvin, Rimita Lahiri, Aditya Kommineni et al.

Reliable transcription of child-adult conversations in clinical settings is crucial for diagnosing developmental disorders like Autism. Recent advances in deep learning and availability of large scale transcribed data has led to development of speech foundation models that have shown dramatic improvements in ASR performance. However, their performance on conversational child-adult interactions remains underexplored. In this work, we provide a comprehensive evaluation of ASR performance on a dataset containing child-adult interactions from autism diagnostic sessions, using Whisper, Wav2Vec2, HuBERT, and WavLM. We find that speech foundation models show a noticeable performance drop (15-20% absolute WER) for child speech compared to adult speech in the conversational setting. Then, we fine-tune the best-performing zero-shot model (Whisper-large) using LoRA in a low-resource setting, yielding 8% and 13% absolute WER improvements for child and adult speech, respectively.

Shih-Heng Wang, Tiantian Feng, Aditya Kommineni et al.

Neural Audio Codecs (NACs) are widely adopted in modern speech systems, yet how they encode linguistic and paralinguistic information remains unclear. Improving the interpretability of NAC representations is critical for understanding and deploying them in sensitive applications. Hence, we employ Sparse Autoencoders (SAEs) to decompose dense NAC representations into sparse, interpretable activations. In this work, we focus on a challenging paralinguistic attribute-accent-and propose a framework to quantify NAC interpretability. We evaluate four NAC models under 16 SAE configurations using a relative performance index. Our results show that DAC and SpeechTokenizer achieve the highest interpretability. We further reveal that acoustic-oriented NACs encode accent information primarily in activation magnitudes of sparse representations, whereas phonetic-oriented NACs rely more on activation positions, and that low-bitrate EnCodec variants show higher interpretability.

Simon Pistrosch, Kleanthis Avramidis, Tiantian Feng et al.

The expression of affect is integral to spoken communication, yet, its link to underlying articulatory execution remains unclear. Measures of articulatory muscle activity such as EMG could reveal how speech production is modulated by emotion alongside acoustic speech analyses. We investigate affect decoding from facial and neck surface electromyography (sEMG) during phonated and silent speech production. For this purpose, we introduce a dataset comprising 2,780 utterances from 12 participants across 3 tasks, on which we evaluate both intra- and inter-subject decoding using a range of features and model embeddings. Our results reveal that EMG representations reliably discriminate frustration with up to 0.845 AUC, and generalize well across articulation modes. Our ablation study further demonstrates that affective signatures are embedded in facial motor activity and persist in the absence of phonation, highlighting the potential of EMG sensing for affect-aware silent speech interfaces.

Tiantian Feng, Kleanthis Avramidis, Anfeng Xu et al.

Healthcare professionals work in complex, high-stakes environments where effective communication is critical for care delivery, team coordination, and individual well-being. However, communication activity in everyday clinical settings remains challenging to measure and largely unexplored in human behavioral research. We present VoxCare, a scalable egocentric wearable audio sensing and computing system that captures natural communication behaviors of hospital professionals in real-world settings without storing raw audio. VoxCare performs real-time, on-device acoustic feature extraction and applies a speech foundation model-guided teacher-student framework to identify foreground speech activity. From these features, VoxCare derives interpretable behavioral measures of communication frequency, duration, and vocal arousal. Our analyses reveal how, when, and how often clinicians communicate across different shifts and working units, and suggest that communication activity reflects underlying workload and stress. By enabling continuous assessment of communication patterns in everyday contexts, this study provides data-driven approaches to understand the behaviors of healthcare providers and ultimately improve healthcare delivery.

Emily Zhou, Marcus Ma, Kleanthis Avramidis et al.

Eye movement and memory retrieval are deeply and bidirectionally intertwined, however existing literature is generally confined to controlled lab settings. We investigate the relationship between eye gaze and memory recall in free-form autobiographical recall, which comprises both autonoetic consciousness -- the ability to mentally place oneself in the past or future -- and various affective states. Using a large video corpus of semi-naturalistic interviews with Holocaust survivors (N = 806), we examine eye movements with respect to episodic, semantic, affective, and temporal dimensions of traumatic and highly emotional autobiographical recall. We observe gaze patterns vary significantly across certain temporal contexts, most prominently in vertical eye movements. We additionally train intra-subject sequence models to predict temporal context of sentences from segments of gaze features, and find that eye movements entirely preceding sentence onset are sufficient for prediction. Our results corroborate prior findings in literature linking eye movements to memory in controlled and semi-structured settings, reinforcing the role of eye gaze in retrieving and constructing memories, especially in highly emotional and remote memory recall.

Xuan Shi, Chang Zeng, Tiantian Feng et al.

Speech tokenizers are essential for connecting speech to large language models (LLMs) in multimodal systems. These tokenizers are expected to preserve both semantic and acoustic information for downstream understanding and generation. However, emerging evidence suggests that what is termed "semantic" in speech representations does not align with text-derived semantics: a mismatch that can degrade multimodal LLM performance. In this paper, we systematically analyze the information encoded by several widely used speech tokenizers, disentangling their semantic and phonetic content through word-level probing tasks, layerwise representation analysis, and cross-modal alignment metrics such as CKA. Our results show that current tokenizers primarily capture phonetic rather than lexical-semantic structure, and we derive practical implications for the design of next-generation speech tokenization methods.

Zhonghao Shi, Xuan Shi, Anfeng Xu et al.

Automatic speech recognition (ASR) models often experience performance degradation due to data domain shifts introduced at test time, a challenge that is further amplified for child speakers. Test-time adaptation (TTA) methods have shown great potential in bridging this domain gap. However, the use of TTA to adapt ASR models to the individual differences in each child's speech has not yet been systematically studied. In this work, we investigate the effectiveness of two widely used TTA methods-SUTA, SGEM-in adapting off-the-shelf ASR models and their fine-tuned versions for child speech recognition, with the goal of enabling continuous, unsupervised adaptation at test time. Our findings show that TTA significantly improves the performance of both off-the-shelf and fine-tuned ASR models, both on average and across individual child speakers, compared to unadapted baselines. However, while TTA helps adapt to individual variability, it may still be limited with non-linguistic child speech.

Angelly Cabrera, Kleanthis Avramidis, Shrikanth Narayanan

Coffee leaf rust, a foliar disease caused by the fungus Hemileia vastatrix, poses a major threat to coffee production, especially in Central America. Climate change further aggravates this issue, as it shortens the latency period between initial infection and the emergence of visible symptoms in diseases like leaf rust. Shortened latency periods can lead to more severe plant epidemics and faster spread of diseases. There is, hence, an urgent need for effective disease management strategies. To address these challenges, we explore the potential of deep learning models for enhancing early disease detection. However, deep learning models require extensive processing power and large amounts of data for model training, resources that are typically scarce. To overcome these barriers, we propose a preprocessing technique that involves convolving training images with a high-pass filter to enhance lesion-leaf contrast, significantly improving model efficacy in resource-limited environments. This method and our model demonstrated a strong performance, achieving over 90% across all evaluation metrics--including precision, recall, F1-score, and the Dice coefficient. Our experiments show that this approach outperforms other methods, including two different image preprocessing techniques and using unaltered, full-color images.

Kai-Wei Chang, Yi-Cheng Lin, Huang-Cheng Chou et al.

Speech technologies have advanced rapidly and serve diverse populations worldwide. However, many languages remain underrepresented due to limited resources. In this paper, we introduce \textbf{TaigiSpeech}, a real-world speech intent dataset in Taiwanese Taigi (aka Taiwanese Hokkien/Southern Min), which is a low-resource and primarily spoken language. The dataset is collected from older adults, comprising 21 speakers with a total of 3k utterances. It is designed for practical intent detection scenarios, including healthcare and home assistant applications. To address the scarcity of labeled data, we explore two data mining strategies with two levels of supervision: keyword match data mining with LLM pseudo labeling via an intermediate language and an audio-visual framework that leverages multimodal cues with minimal textual supervision. This design enables scalable dataset construction for low-resource and unwritten spoken languages. TaigiSpeech will be released under the CC BY 4.0 license to facilitate broad adoption and research on low-resource and unwritten languages. The project website and the dataset can be found on https://kwchang.org/taigispeech.

Daniel Yang, Aditya Kommineni, Mohammad Alshehri et al.

The lack of contextual information in text data can make the annotation process of text-based emotion classification datasets challenging. As a result, such datasets often contain labels that fail to consider all the relevant emotions in the vocabulary. This misalignment between text inputs and labels can degrade the performance of machine learning models trained on top of them. As re-annotating entire datasets is a costly and time-consuming task that cannot be done at scale, we propose to use the expressive capabilities of large language models to synthesize additional context for input text to increase its alignment with the annotated emotional labels. In this work, we propose a formal definition of textual context to motivate a prompting strategy to enhance such contextual information. We provide both human and empirical evaluation to demonstrate the efficacy of the enhanced context. Our method improves alignment between inputs and their human-annotated labels from both an empirical and human-evaluated standpoint.

Chenxiao Yu, Bowen Yi, Farzan Karimi-Malekabadi et al.

Large language models (LLMs) often produce human-like moral judgments, but it is unclear whether this reflects an internal conceptual structure or superficial ``moral mimicry.'' Using Moral Foundations Theory (MFT) as an analytic framework, we study how moral foundations are encoded, organized, and expressed within two instruction-tuned LLMs: Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct. We employ a multi-level approach combining (i) layer-wise analysis of MFT concept representations and their alignment with human moral perceptions, (ii) pretrained sparse autoencoders (SAEs) over the residual stream to identify sparse features that support moral concepts, and (iii) causal steering interventions using dense MFT vectors and sparse SAE features. We find that both models represent and distinguish moral foundations in a structured, layer-dependent way that aligns with human judgments. At a finer scale, SAE features show clear semantic links to specific foundations, suggesting partially disentangled mechanisms within shared representations. Finally, steering along either dense vectors or sparse features produces predictable shifts in foundation-relevant behavior, demonstrating a causal connection between internal representations and moral outputs. Together, our results provide mechanistic evidence that moral concepts in LLMs are distributed, layered, and partly disentangled, suggesting that pluralistic moral structure can emerge as a latent pattern from the statistical regularities of language alone.

Tiantian Feng, Kevin Huang, Anfeng Xu et al.

We present Voxlect, a novel benchmark for modeling dialects and regional languages worldwide using speech foundation models. Specifically, we report comprehensive benchmark evaluations on dialects and regional language varieties in English, Arabic, Mandarin and Cantonese, Tibetan, Indic languages, Thai, Spanish, French, German, Brazilian Portuguese, and Italian. Our study used over 2 million training utterances from 30 publicly available speech corpora that are provided with dialectal information. We evaluate the performance of several widely used speech foundation models in classifying speech dialects. We assess the robustness of the dialectal models under noisy conditions and present an error analysis that highlights modeling results aligned with geographic continuity. In addition to benchmarking dialect classification, we demonstrate several downstream applications enabled by Voxlect. Specifically, we show that Voxlect can be applied to augment existing speech recognition datasets with dialect information, enabling a more detailed analysis of ASR performance across dialectal variations. Voxlect is also used as a tool to evaluate the performance of speech generation systems. Voxlect is publicly available with the license of the RAIL family at: https://github.com/tiantiaf0627/voxlect.

Georgios Chochlakis, Peter Wu, Arjun Bedi et al.

Modeling complex subjective tasks in Natural Language Processing, such as recognizing emotion and morality, is considerably challenging due to significant variation in human annotations. This variation often reflects reasonable differences in semantic interpretations rather than mere noise, necessitating methods to distinguish between legitimate subjectivity and error. We address this challenge by exploring label verification in these contexts using Large Language Models (LLMs). First, we propose a simple In-Context Learning binary filtering baseline that estimates the reasonableness of a document-label pair. We then introduce the Label-in-a-Haystack setting: the query and its label(s) are included in the demonstrations shown to LLMs, which are prompted to predict the label(s) again, while receiving task-specific instructions (e.g., emotion recognition) rather than label copying. We show how the failure to copy the label(s) to the output of the LLM are task-relevant and informative. Building on this, we propose the Label-in-a-Haystack Rectification (LiaHR) framework for subjective label correction: when the model outputs diverge from the reference gold labels, we assign the generated labels to the example instead of discarding it. This approach can be integrated into annotation pipelines to enhance signal-to-noise ratios. Comprehensive analyses, human evaluations, and ecological validity studies verify the utility of LiaHR for label correction. Code is available at https://github.com/gchochla/liahr.

Tiantian Feng, Dimitrios Dimitriadis, Shrikanth Narayanan

Recent advances in foundation models have enabled audio-generative models that produce high-fidelity sounds associated with music, events, and human actions. Despite the success achieved in modern audio-generative models, the conventional approach to assessing the quality of the audio generation relies heavily on distance metrics like Frechet Audio Distance. In contrast, we aim to evaluate the quality of audio generation by examining the effectiveness of using them as training data. Specifically, we conduct studies to explore the use of synthetic audio for audio recognition. Moreover, we investigate whether synthetic audio can serve as a resource for data augmentation in speech-related modeling. Our comprehensive experiments demonstrate the potential of using synthetic audio for audio recognition and speech-related modeling. Our code is available at https://github.com/usc-sail/SynthAudio.

Jihwan Lee, Aditya Kommineni, Tiantian Feng et al.

Speech decoding from EEG signals is a challenging task, where brain activity is modeled to estimate salient characteristics of acoustic stimuli. We propose FESDE, a novel framework for Fully-End-to-end Speech Decoding from EEG signals. Our approach aims to directly reconstruct listened speech waveforms given EEG signals, where no intermediate acoustic feature processing step is required. The proposed method consists of an EEG module and a speech module along with a connector. The EEG module learns to better represent EEG signals, while the speech module generates speech waveforms from model representations. The connector learns to bridge the distributions of the latent spaces of EEG and speech. The proposed framework is both simple and efficient, by allowing single-step inference, and outperforms prior works on objective metrics. A fine-grained phoneme analysis is conducted to unveil model characteristics of speech decoding. The source code is available here: github.com/lee-jhwn/fesde.

Manoj Kumar, Tae Jin-Park, Somer Bishop et al.

Neural speaker embeddings trained using classification objectives have demonstrated state-of-the-art performance in multiple applications. Typically, such embeddings are trained on an out-of-domain corpus on a single task e.g., speaker classification, albeit with a large number of classes (speakers). In this work, we reformulate embedding training under the meta-learning paradigm. We redistribute the training corpus as an ensemble of multiple related speaker classification tasks, and learn a representation that generalizes better to unseen speakers. First, we develop an open source toolkit to train x-vectors that is matched in performance with pre-trained Kaldi models for speaker diarization and speaker verification applications. We find that different bottleneck layers in the architecture variedly favor different applications. Next, we use two meta-learning strategies, namely prototypical networks and relation networks, to improve over the x-vector embeddings. Our best performing model achieves a relative improvement of 12.37% and 7.11% in speaker error on the DIHARD II development corpus and the AMI meeting corpus, respectively. We analyze improvements across different domains in the DIHARD corpus. Notably, on the challenging child speech domain, we study the relation between child age and the diarization performance. Further, we show reductions in equal error rate for speaker verification on the SITW corpus (7.68%) and the VOiCES challenge corpus (8.78%). We observe that meta-learning particularly offers benefits in challenging acoustic conditions and recording setups encountered in these corpora. Our experiments illustrate the applicability of meta-learning as a generalized learning paradigm for training deep neural speaker embeddings.

Jordan Prescott, Thanathai Lertpetchpun, Shrikanth Narayanan

Adversarial perturbations exploit vulnerabilities in automatic speech recognition (ASR) systems while preserving human perceived linguistic content. Neural audio codecs impose a discrete bottleneck that can suppress fine-grained signal variations associated with adversarial noise. We examine how the granularity of this bottleneck, controlled by residual vector quantization (RVQ) depth, shapes adversarial robustness. We observe a non-monotonic trade-off under gradient-based attacks: shallow quantization suppresses adversarial perturbations but degrades speech content, while deeper quantization preserves both content and perturbations. Intermediate depths balance these effects and minimize transcription error. We further show that adversarially induced changes in discrete codebook tokens strongly correlate with transcription error. These gains persist under adaptive attacks, where neural codec configurations outperform traditional compression defenses.

Girish Narayanswamy, Xin Liu, Kumar Ayush et al.

Wearable sensors have become ubiquitous thanks to a variety of health tracking features. The resulting continuous and longitudinal measurements from everyday life generate large volumes of data; however, making sense of these observations for scientific and actionable insights is non-trivial. Inspired by the empirical success of generative modeling, where large neural networks learn powerful representations from vast amounts of text, image, video, or audio data, we investigate the scaling properties of sensor foundation models across compute, data, and model size. Using a dataset of up to 40 million hours of in-situ heart rate, heart rate variability, electrodermal activity, accelerometer, skin temperature, and altimeter per-minute data from over 165,000 people, we create LSM, a multimodal foundation model built on the largest wearable-signals dataset with the most extensive range of sensor modalities to date. Our results establish the scaling laws of LSM for tasks such as imputation, interpolation and extrapolation, both across time and sensor modalities. Moreover, we highlight how LSM enables sample-efficient downstream learning for tasks like exercise and activity recognition.