Björn W. Schuller

Joseph Turian, Jordie Shier, Humair Raj Khan et al. · cmu

What audio embedding approach generalizes best to a wide range of downstream tasks across a variety of everyday domains without fine-tuning? The aim of the HEAR benchmark is to develop a general-purpose audio representation that provides a strong basis for learning in a wide variety of tasks and scenarios. HEAR evaluates audio representations using a benchmark suite across a variety of domains, including speech, environmental sound, and music. HEAR was launched as a NeurIPS 2021 shared challenge. In the spirit of shared exchange, each participant submitted an audio embedding model following a common API that is general-purpose, open-source, and freely available to use. Twenty-nine models by thirteen external teams were evaluated on nineteen diverse downstream tasks derived from sixteen datasets. Open evaluation code, submitted models and datasets are key contributions, enabling comprehensive and reproducible evaluation, as well as previously impossible longitudinal studies. It still remains an open question whether one single general-purpose audio representation can perform as holistically as the human ear.

Andreas Triantafyllopoulos, Johannes Wagner, Hagen Wierstorf et al.

Large, pre-trained neural networks consisting of self-attention layers (transformers) have recently achieved state-of-the-art results on several speech emotion recognition (SER) datasets. These models are typically pre-trained in self-supervised manner with the goal to improve automatic speech recognition performance -- and thus, to understand linguistic information. In this work, we investigate the extent in which this information is exploited during SER fine-tuning. Using a reproducible methodology based on open-source tools, we synthesise prosodically neutral speech utterances while varying the sentiment of the text. Valence predictions of the transformer model are very reactive to positive and negative sentiment content, as well as negations, but not to intensifiers or reducers, while none of those linguistic features impact arousal or dominance. These findings show that transformers can successfully leverage linguistic information to improve their valence predictions, and that linguistic analysis should be included in their testing.

Lukas Christ, Shahin Amiriparian, Alexander Kathan et al.

Humor is a substantial element of human social behavior, affect, and cognition. Its automatic understanding can facilitate a more naturalistic human-AI interaction. Current methods of humor detection have been exclusively based on staged data, making them inadequate for "real-world" applications. We contribute to addressing this deficiency by introducing the novel Passau-Spontaneous Football Coach Humor (Passau-SFCH) dataset, comprising about 11 hours of recordings. The Passau-SFCH dataset is annotated for the presence of humor and its dimensions (sentiment and direction) as proposed in Martin's Humor Style Questionnaire. We conduct a series of experiments employing pretrained Transformers, convolutional neural networks, and expert-designed features. The performance of each modality (text, audio, video) for spontaneous humor recognition is analyzed and their complementarity is investigated. Our findings suggest that for the automatic analysis of humor and its sentiment, facial expressions are most promising, while humor direction can be best modeled via text-based features. Further, we experiment with different multimodal approaches to humor recognition, including decision-level fusion and MulT, a multimodal Transformer approach. In this context, we propose a novel multimodal architecture that yields the best overall results. Finally, we make our code publicly available at https://www.github.com/lc0197/passau-sfch. The Passau-SFCH dataset is available upon request.

Lukas Christ, Shahin Amiriparian, Manuel Milling et al.

Telling stories is an integral part of human communication which can evoke emotions and influence the affective states of the audience. Automatically modelling emotional trajectories in stories has thus attracted considerable scholarly interest. However, as most existing works have been limited to unsupervised dictionary-based approaches, there is no labelled benchmark for this task. We address this gap by introducing continuous valence and arousal annotations for an existing dataset of children's stories annotated with discrete emotion categories. We collect additional annotations for this data and map the originally categorical labels to the valence and arousal space. Leveraging recent advances in Natural Language Processing, we propose a set of novel Transformer-based methods for predicting valence and arousal signals over the course of written stories. We explore several strategies for fine-tuning a pretrained ELECTRA model and study the benefits of considering a sentence's context when inferring its emotionality. Moreover, we experiment with additional LSTM and Transformer layers. The best configuration achieves a Concordance Correlation Coefficient (CCC) of .7338 for valence and .6302 for arousal on the test set, demonstrating the suitability of our proposed approach. Our code and additional annotations are made available at https://github.com/lc0197/emotion_modelling_stories.

Johannes Wagner, Andreas Triantafyllopoulos, Hagen Wierstorf et al.

Recent advances in transformer-based architectures which are pre-trained in self-supervised manner have shown great promise in several machine learning tasks. In the audio domain, such architectures have also been successfully utilised in the field of speech emotion recognition (SER). However, existing works have not evaluated the influence of model size and pre-training data on downstream performance, and have shown limited attention to generalisation, robustness, fairness, and efficiency. The present contribution conducts a thorough analysis of these aspects on several pre-trained variants of wav2vec 2.0 and HuBERT that we fine-tuned on the dimensions arousal, dominance, and valence of MSP-Podcast, while additionally using IEMOCAP and MOSI to test cross-corpus generalisation. To the best of our knowledge, we obtain the top performance for valence prediction without use of explicit linguistic information, with a concordance correlation coefficient (CCC) of .638 on MSP-Podcast. Furthermore, our investigations reveal that transformer-based architectures are more robust to small perturbations compared to a CNN-based baseline and fair with respect to biological sex groups, but not towards individual speakers. Finally, we are the first to show that their extraordinary success on valence is based on implicit linguistic information learnt during fine-tuning of the transformer layers, which explains why they perform on-par with recent multimodal approaches that explicitly utilise textual information. Our findings collectively paint the following picture: transformer-based architectures constitute the new state-of-the-art in SER, but further advances are needed to mitigate remaining robustness and individual speaker issues. To make our findings reproducible, we release the best performing model to the community.

Zhao Ren, Yi Chang, Thanh Tam Nguyen et al.

Heart sound auscultation has been applied in clinical usage for early screening of cardiovascular diseases. Due to the high demand for auscultation expertise, automatic auscultation can help with auxiliary diagnosis and reduce the burden of training professional clinicians. Nevertheless, there is a limit to classic machine learning's performance improvement in the era of big data. Deep learning has outperformed classic machine learning in many research fields, as it employs more complex model architectures with a stronger capability of extracting effective representations. Moreover, it has been successfully applied to heart sound analysis in the past years. As most review works about heart sound analysis were carried out before 2017, the present survey is the first to work on a comprehensive overview to summarise papers on heart sound analysis with deep learning published in 2017--2022. This work introduces both classic machine learning and deep learning for comparison, and further offer insights about the advances and future research directions in deep learning for heart sound analysis. Our repository is publicly available at \url{https://github.com/zhaoren91/awesome-heart-sound-analysis}.

Zheng Lian, Haiyang Sun, Licai Sun et al.

The first Multimodal Emotion Recognition Challenge (MER 2023) was successfully held at ACM Multimedia. The challenge focuses on system robustness and consists of three distinct tracks: (1) MER-MULTI, where participants are required to recognize both discrete and dimensional emotions; (2) MER-NOISE, in which noise is added to test videos for modality robustness evaluation; (3) MER-SEMI, which provides a large amount of unlabeled samples for semi-supervised learning. In this paper, we introduce the motivation behind this challenge, describe the benchmark dataset, and provide some statistics about participants. To continue using this dataset after MER 2023, please sign a new End User License Agreement and send it to our official email address merchallenge.contact@gmail.com. We believe this high-quality dataset can become a new benchmark in multimodal emotion recognition, especially for the Chinese research community.

Andreas Triantafyllopoulos, Björn W. Schuller, Gökçe İymen et al.

Speech is the fundamental mode of human communication, and its synthesis has long been a core priority in human-computer interaction research. In recent years, machines have managed to master the art of generating speech that is understandable by humans. But the linguistic content of an utterance encompasses only a part of its meaning. Affect, or expressivity, has the capacity to turn speech into a medium capable of conveying intimate thoughts, feelings, and emotions -- aspects that are essential for engaging and naturalistic interpersonal communication. While the goal of imparting expressivity to synthesised utterances has so far remained elusive, following recent advances in text-to-speech synthesis, a paradigm shift is well under way in the fields of affective speech synthesis and conversion as well. Deep learning, as the technology which underlies most of the recent advances in artificial intelligence, is spearheading these efforts. In the present overview, we outline ongoing trends and summarise state-of-the-art approaches in an attempt to provide a comprehensive overview of this exciting field.

Mostafa M. Amin, Erik Cambria, Björn W. Schuller

ChatGPT has shown the potential of emerging general artificial intelligence capabilities, as it has demonstrated competent performance across many natural language processing tasks. In this work, we evaluate the capabilities of ChatGPT to perform text classification on three affective computing problems, namely, big-five personality prediction, sentiment analysis, and suicide tendency detection. We utilise three baselines, a robust language model (RoBERTa-base), a legacy word model with pretrained embeddings (Word2Vec), and a simple bag-of-words baseline (BoW). Results show that the RoBERTa trained for a specific downstream task generally has a superior performance. On the other hand, ChatGPT provides decent results, and is relatively comparable to the Word2Vec and BoW baselines. ChatGPT further shows robustness against noisy data, where Word2Vec models achieve worse results due to noise. Results indicate that ChatGPT is a good generalist model that is capable of achieving good results across various problems without any specialised training, however, it is not as good as a specialised model for a downstream task.

Lukas Christ, Shahin Amiriparian, Alice Baird et al.

The Multimodal Sentiment Analysis Challenge (MuSe) 2022 is dedicated to multimodal sentiment and emotion recognition. For this year's challenge, we feature three datasets: (i) the Passau Spontaneous Football Coach Humor (Passau-SFCH) dataset that contains audio-visual recordings of German football coaches, labelled for the presence of humour; (ii) the Hume-Reaction dataset in which reactions of individuals to emotional stimuli have been annotated with respect to seven emotional expression intensities, and (iii) the Ulm-Trier Social Stress Test (Ulm-TSST) dataset comprising of audio-visual data labelled with continuous emotion values (arousal and valence) of people in stressful dispositions. Using the introduced datasets, MuSe 2022 2022 addresses three contemporary affective computing problems: in the Humor Detection Sub-Challenge (MuSe-Humor), spontaneous humour has to be recognised; in the Emotional Reactions Sub-Challenge (MuSe-Reaction), seven fine-grained `in-the-wild' emotions have to be predicted; and in the Emotional Stress Sub-Challenge (MuSe-Stress), a continuous prediction of stressed emotion values is featured. The challenge is designed to attract different research communities, encouraging a fusion of their disciplines. Mainly, MuSe 2022 targets the communities of audio-visual emotion recognition, health informatics, and symbolic sentiment analysis. This baseline paper describes the datasets as well as the feature sets extracted from them. A recurrent neural network with LSTM cells is used to set competitive baseline results on the test partitions for each sub-challenge. We report an Area Under the Curve (AUC) of .8480 for MuSe-Humor; .2801 mean (from 7-classes) Pearson's Correlations Coefficient for MuSe-Reaction, as well as .4931 Concordance Correlation Coefficient (CCC) and .4761 for valence and arousal in MuSe-Stress, respectively.

Mostafa M. Amin, Rui Mao, Erik Cambria et al.

With the rise of foundation models, a new artificial intelligence paradigm has emerged, by simply using general purpose foundation models with prompting to solve problems instead of training a separate machine learning model for each problem. Such models have been shown to have emergent properties of solving problems that they were not initially trained on. The studies for the effectiveness of such models are still quite limited. In this work, we widely study the capabilities of the ChatGPT models, namely GPT-4 and GPT-3.5, on 13 affective computing problems, namely aspect extraction, aspect polarity classification, opinion extraction, sentiment analysis, sentiment intensity ranking, emotions intensity ranking, suicide tendency detection, toxicity detection, well-being assessment, engagement measurement, personality assessment, sarcasm detection, and subjectivity detection. We introduce a framework to evaluate the ChatGPT models on regression-based problems, such as intensity ranking problems, by modelling them as pairwise ranking classification. We compare ChatGPT against more traditional NLP methods, such as end-to-end recurrent neural networks and transformers. The results demonstrate the emergent abilities of the ChatGPT models on a wide range of affective computing problems, where GPT-3.5 and especially GPT-4 have shown strong performance on many problems, particularly the ones related to sentiment, emotions, or toxicity. The ChatGPT models fell short for problems with implicit signals, such as engagement measurement and subjectivity detection.

Mani Kumar Tellamekala, Shahin Amiriparian, Björn W. Schuller et al.

Automatically recognising apparent emotions from face and voice is hard, in part because of various sources of uncertainty, including in the input data and the labels used in a machine learning framework. This paper introduces an uncertainty-aware audiovisual fusion approach that quantifies modality-wise uncertainty towards emotion prediction. To this end, we propose a novel fusion framework in which we first learn latent distributions over audiovisual temporal context vectors separately, and then constrain the variance vectors of unimodal latent distributions so that they represent the amount of information each modality provides w.r.t. emotion recognition. In particular, we impose Calibration and Ordinal Ranking constraints on the variance vectors of audiovisual latent distributions. When well-calibrated, modality-wise uncertainty scores indicate how much their corresponding predictions may differ from the ground truth labels. Well-ranked uncertainty scores allow the ordinal ranking of different frames across the modalities. To jointly impose both these constraints, we propose a softmax distributional matching loss. In both classification and regression settings, we compare our uncertainty-aware fusion model with standard model-agnostic fusion baselines. Our evaluation on two emotion recognition corpora, AVEC 2019 CES and IEMOCAP, shows that audiovisual emotion recognition can considerably benefit from well-calibrated and well-ranked latent uncertainty measures.

Rodrigo Mira, Alexandros Haliassos, Stavros Petridis et al.

Video-to-speech synthesis (also known as lip-to-speech) refers to the translation of silent lip movements into the corresponding audio. This task has received an increasing amount of attention due to its self-supervised nature (i.e., can be trained without manual labelling) combined with the ever-growing collection of audio-visual data available online. Despite these strong motivations, contemporary video-to-speech works focus mainly on small- to medium-sized corpora with substantial constraints in both vocabulary and setting. In this work, we introduce a scalable video-to-speech framework consisting of two components: a video-to-spectrogram predictor and a pre-trained neural vocoder, which converts the mel-frequency spectrograms into waveform audio. We achieve state-of-the art results for GRID and considerably outperform previous approaches on LRW. More importantly, by focusing on spectrogram prediction using a simple feedforward model, we can efficiently and effectively scale our method to very large and unconstrained datasets: To the best of our knowledge, we are the first to show intelligible results on the challenging LRS3 dataset.

Björn W. Schuller, Anton Batliner, Shahin Amiriparian et al.

The ACM Multimedia 2022 Computational Paralinguistics Challenge addresses four different problems for the first time in a research competition under well-defined conditions: In the Vocalisations and Stuttering Sub-Challenges, a classification on human non-verbal vocalisations and speech has to be made; the Activity Sub-Challenge aims at beyond-audio human activity recognition from smartwatch sensor data; and in the Mosquitoes Sub-Challenge, mosquitoes need to be detected. We describe the Sub-Challenges, baseline feature extraction, and classifiers based on the usual ComPaRE and BoAW features, the auDeep toolkit, and deep feature extraction from pre-trained CNNs using the DeepSpectRum toolkit; in addition, we add end-to-end sequential modelling, and a log-mel-128-BNN.

Yuezhou Zhang, Amos A Folarin, Judith Dineley et al.

Language use has been shown to correlate with depression, but large-scale validation is needed. Traditional methods like clinic studies are expensive. So, natural language processing has been employed on social media to predict depression, but limitations remain-lack of validated labels, biased user samples, and no context. Our study identified 29 topics in 3919 smartphone-collected speech recordings from 265 participants using the Whisper tool and BERTopic model. Six topics with a median PHQ-8 greater than or equal to 10 were regarded as risk topics for depression: No Expectations, Sleep, Mental Therapy, Haircut, Studying, and Coursework. To elucidate the topic emergence and associations with depression, we compared behavioral (from wearables) and linguistic characteristics across identified topics. The correlation between topic shifts and changes in depression severity over time was also investigated, indicating the importance of longitudinally monitoring language use. We also tested the BERTopic model on a similar smaller dataset (356 speech recordings from 57 participants), obtaining some consistent results. In summary, our findings demonstrate specific speech topics may indicate depression severity. The presented data-driven workflow provides a practical approach to collecting and analyzing large-scale speech data from real-world settings for digital health research.

Björn W. Schuller, Anton Batliner, Shahin Amiriparian et al.

The ACM Multimedia 2023 Computational Paralinguistics Challenge addresses two different problems for the first time in a research competition under well-defined conditions: In the Emotion Share Sub-Challenge, a regression on speech has to be made; and in the Requests Sub-Challenges, requests and complaints need to be detected. We describe the Sub-Challenges, baseline feature extraction, and classifiers based on the usual ComPaRE features, the auDeep toolkit, and deep feature extraction from pre-trained CNNs using the DeepSpectRum toolkit; in addition, wav2vec2 models are used.

Mani Kumar Tellamekala, Ömer Sümer, Björn W. Schuller et al.

Recognising continuous emotions and action unit (AU) intensities from face videos requires a spatial and temporal understanding of expression dynamics. Existing works primarily rely on 2D face appearances to extract such dynamics. This work focuses on a promising alternative based on parametric 3D face shape alignment models, which disentangle different factors of variation, including expression-induced shape variations. We aim to understand how expressive 3D face shapes are in estimating valence-arousal and AU intensities compared to the state-of-the-art 2D appearance-based models. We benchmark four recent 3D face alignment models: ExpNet, 3DDFA-V2, DECA, and EMOCA. In valence-arousal estimation, expression features of 3D face models consistently surpassed previous works and yielded an average concordance correlation of .739 and .574 on SEWA and AVEC 2019 CES corpora, respectively. We also study how 3D face shapes performed on AU intensity estimation on BP4D and DISFA datasets, and report that 3D face features were on par with 2D appearance features in AUs 4, 6, 10, 12, and 25, but not the entire set of AUs. To understand this discrepancy, we conduct a correspondence analysis between valence-arousal and AUs, which points out that accurate prediction of valence-arousal may require the knowledge of only a few AUs.

Vincent Karas, Mani Kumar Tellamekala, Adria Mallol-Ragolta et al.

In this paper, we present our submission to 3rd Affective Behavior Analysis in-the-wild (ABAW) challenge. Learningcomplex interactions among multimodal sequences is critical to recognise dimensional affect from in-the-wild audiovisual data. Recurrence and attention are the two widely used sequence modelling mechanisms in the literature. To clearly understand the performance differences between recurrent and attention models in audiovisual affect recognition, we present a comprehensive evaluation of fusion models based on LSTM-RNNs, self-attention and cross-modal attention, trained for valence and arousal estimation. Particularly, we study the impact of some key design choices: the modelling complexity of CNN backbones that provide features to the the temporal models, with and without end-to-end learning. We trained the audiovisual affect recognition models on in-the-wild ABAW corpus by systematically tuning the hyper-parameters involved in the network architecture design and training optimisation. Our extensive evaluation of the audiovisual fusion models shows that LSTM-RNNs can outperform the attention models when coupled with low-complex CNN backbones and trained in an end-to-end fashion, implying that attention models may not necessarily be the optimal choice for continuous-time multimodal emotion recognition.

Mostafa M. Amin, Erik Cambria, Björn W. Schuller

The employment of foundation models is steadily expanding, especially with the launch of ChatGPT and the release of other foundation models. These models have shown the potential of emerging capabilities to solve problems, without being particularly trained to solve. A previous work demonstrated these emerging capabilities in affective computing tasks; the performance quality was similar to traditional Natural Language Processing (NLP) techniques, but falling short of specialised trained models, like fine-tuning of the RoBERTa language model. In this work, we extend this by exploring if ChatGPT has novel knowledge that would enhance existing specialised models when they are fused together. We achieve this by investigating the utility of verbose responses from ChatGPT about solving a downstream task, in addition to studying the utility of fusing that with existing NLP methods. The study is conducted on three affective computing problems, namely sentiment analysis, suicide tendency detection, and big-five personality assessment. The results conclude that ChatGPT has indeed novel knowledge that can improve existing NLP techniques by way of fusion, be it early or late fusion.

Mohammad Nadeem, Shahab Saquib Sohail, Erik Cambria et al.

Foundational Large Language Models (LLMs) have changed the way we perceive technology. They have been shown to excel in tasks ranging from poem writing and coding to essay generation and puzzle solving. With the incorporation of image generation capability, they have become more comprehensive and versatile AI tools. At the same time, researchers are striving to identify the limitations of these tools to improve them further. Currently identified flaws include hallucination, biases, and bypassing restricted commands to generate harmful content. In the present work, we have identified a fundamental limitation related to the image generation ability of LLMs, and termed it The NO Syndrome. This negation blindness refers to LLMs inability to correctly comprehend NO related natural language prompts to generate the desired images. Interestingly, all tested LLMs including GPT-4, Gemini, and Copilot were found to be suffering from this syndrome. To demonstrate the generalization of this limitation, we carried out simulation experiments and conducted entropy-based and benchmark statistical analysis tests on various LLMs in multiple languages, including English, Hindi, and French. We conclude that the NO syndrome is a significant flaw in current LLMs that needs to be addressed. A related finding of this study showed a consistent discrepancy between image and textual responses as a result of this NO syndrome. We posit that the introduction of a negation context-aware reinforcement learning based feedback loop between the LLMs textual response and generated image could help ensure the generated text is based on both the LLMs correct contextual understanding of the negation query and the generated visual output.

Yi Chang, Zhao Ren, Thanh Tam Nguyen et al.

Speech emotion recognition (SER) has been a popular research topic in human-computer interaction (HCI). As edge devices are rapidly springing up, applying SER to edge devices is promising for a huge number of HCI applications. Although deep learning has been investigated to improve the performance of SER by training complex models, the memory space and computational capability of edge devices represents a constraint for embedding deep learning models. We propose a neural structured learning (NSL) framework through building synthesized graphs. An SER model is trained on a source dataset and used to build graphs on a target dataset. A relatively lightweight model is then trained with the speech samples and graphs together as the input. Our experiments demonstrate that training a lightweight SER model on the target dataset with speech samples and graphs can not only produce small SER models, but also enhance the model performance compared to models with speech samples only and those using classic transfer learning strategies.

Jobie Budd, Kieran Baker, Emma Karoune et al.

The UK COVID-19 Vocal Audio Dataset is designed for the training and evaluation of machine learning models that classify SARS-CoV-2 infection status or associated respiratory symptoms using vocal audio. The UK Health Security Agency recruited voluntary participants through the national Test and Trace programme and the REACT-1 survey in England from March 2021 to March 2022, during dominant transmission of the Alpha and Delta SARS-CoV-2 variants and some Omicron variant sublineages. Audio recordings of volitional coughs, exhalations, and speech were collected in the 'Speak up to help beat coronavirus' digital survey alongside demographic, self-reported symptom and respiratory condition data, and linked to SARS-CoV-2 test results. The UK COVID-19 Vocal Audio Dataset represents the largest collection of SARS-CoV-2 PCR-referenced audio recordings to date. PCR results were linked to 70,794 of 72,999 participants and 24,155 of 25,776 positive cases. Respiratory symptoms were reported by 45.62% of participants. This dataset has additional potential uses for bioacoustics research, with 11.30% participants reporting asthma, and 27.20% with linked influenza PCR test results.

Alexander Gebhard, Andreas Triantafyllopoulos, Teresa Bez et al.

Advances in passive acoustic monitoring and machine learning have led to the procurement of vast datasets for computational bioacoustic research. Nevertheless, data scarcity is still an issue for rare and underrepresented species. This study investigates how meta-information can improve zero-shot audio classification, utilising bird species as an example case study due to the availability of rich and diverse meta-data. We investigate three different sources of metadata: textual bird sound descriptions encoded via (S)BERT, functional traits (AVONET), and bird life-history (BLH) characteristics. As audio features, we extract audio spectrogram transformer (AST) embeddings and project them to the dimension of the auxiliary information by adopting a single linear layer. Then, we employ the dot product as compatibility function and a standard zero-shot learning ranking hinge loss to determine the correct class. The best results are achieved by concatenating the AVONET and BLH features attaining a mean unweighted F1-score of .233 over five different test sets with 8 to 10 classes.

Rui Liu, Haolin Zuo, Zheng Lian et al.

Emotion and Intent Joint Understanding in Multimodal Conversation (MC-EIU) aims to decode the semantic information manifested in a multimodal conversational history, while inferring the emotions and intents simultaneously for the current utterance. MC-EIU is enabling technology for many human-computer interfaces. However, there is a lack of available datasets in terms of annotation, modality, language diversity, and accessibility. In this work, we propose an MC-EIU dataset, which features 7 emotion categories, 9 intent categories, 3 modalities, i.e., textual, acoustic, and visual content, and two languages, i.e., English and Mandarin. Furthermore, it is completely open-source for free access. To our knowledge, MC-EIU is the first comprehensive and rich emotion and intent joint understanding dataset for multimodal conversation. Together with the release of the dataset, we also develop an Emotion and Intent Interaction (EI$^2$) network as a reference system by modeling the deep correlation between emotion and intent in the multimodal conversation. With comparative experiments and ablation studies, we demonstrate the effectiveness of the proposed EI$^2$ method on the MC-EIU dataset. The dataset and codes will be made available at: https://github.com/MC-EIU/MC-EIU.

Björn W. Schuller, Alican Akman, Yi Chang et al.

Among the seventeen Sustainable Development Goals (SDGs) proposed within the 2030 Agenda and adopted by all the United Nations member states, the 13$^{th}$ SDG is a call for action to combat climate change for a better world. In this work, we provide an overview of areas in which audio intelligence -- a powerful but in this context so far hardly considered technology -- can contribute to overcome climate-related challenges. We categorise potential computer audition applications according to the five elements of earth, water, air, fire, and aether, proposed by the ancient Greeks in their five element theory; this categorisation serves as a framework to discuss computer audition in relation to different ecological aspects. Earth and water are concerned with the early detection of environmental changes and, thus, with the protection of humans and animals, as well as the monitoring of land and aquatic organisms. Aerial audio is used to monitor and obtain information about bird and insect populations. Furthermore, acoustic measures can deliver relevant information for the monitoring and forecasting of weather and other meteorological phenomena. The fourth considered element is fire. Due to the burning of fossil fuels, the resulting increase in CO$_2$ emissions and the associated rise in temperature, fire is used as a symbol for man-made climate change and in this context includes the monitoring of noise pollution, machines, as well as the early detection of wildfires. In all these areas, computer audition can help counteract climate change. Aether then corresponds to the technology itself that makes this possible. This work explores these areas and discusses potential applications, while positioning computer audition in relation to methodological alternatives.

Iosif Tsangko, Andreas Triantafyllopoulos, George Margetis et al.

Blind and low-vision (BLV) audiences remain underserved by visual art descriptions, particularly across languages and in museum settings where privacy and intellectual-property constraints may favour small on-premise vision-language models (VLMs). This pilot study investigates curator-guided multilingual art description with Qwen2.5-VL-3B-Instruct for German, Romanian, and Serbian. We construct a parallel BLV-oriented caption corpus from artwork images and metadata, and compare language-specific LoRA adapters with a single multilingual adapter under a fixed backbone and training budget. Evaluation combines automatic lexical and embedding-based metrics with an LLM-as-Judge protocol calibrated against a small Romanian BLV pilot study. Under our pilot setup, language-specific adapters show more stable controllability and visually grounded description quality for Romanian and Serbian, while multilingual adaptation remains competitive in German. We frame these findings as deployment-oriented evidence for small on-premise VLMs, and highlight the need for larger BLV user studies and broader language coverage before drawing general conclusions about multilingual accessibility.

Harry Coppock, George Nicholson, Ivan Kiskin et al.

Recent work has reported that AI classifiers trained on audio recordings can accurately predict severe acute respiratory syndrome coronavirus 2 (SARSCoV2) infection status. Here, we undertake a large scale study of audio-based deep learning classifiers, as part of the UK governments pandemic response. We collect and analyse a dataset of audio recordings from 67,842 individuals with linked metadata, including reverse transcription polymerase chain reaction (PCR) test outcomes, of whom 23,514 tested positive for SARS CoV 2. Subjects were recruited via the UK governments National Health Service Test-and-Trace programme and the REal-time Assessment of Community Transmission (REACT) randomised surveillance survey. In an unadjusted analysis of our dataset AI classifiers predict SARS-CoV-2 infection status with high accuracy (Receiver Operating Characteristic Area Under the Curve (ROCAUC) 0.846 [0.838, 0.854]) consistent with the findings of previous studies. However, after matching on measured confounders, such as age, gender, and self reported symptoms, our classifiers performance is much weaker (ROC-AUC 0.619 [0.594, 0.644]). Upon quantifying the utility of audio based classifiers in practical settings, we find them to be outperformed by simple predictive scores based on user reported symptoms.

Björn W. Schuller, Shahin Amiriparian, Anton Batliner et al.

Charisma is considered as one's ability to attract and potentially also influence others. Clearly, there can be considerable interest from an artificial intelligence's (AI) perspective to provide it with such skill. Beyond, a plethora of use cases opens up for computational measurement of human charisma, such as for tutoring humans in the acquisition of charisma, mediating human-to-human conversation, or identifying charismatic individuals in big social data. A number of models exist that base charisma on various dimensions, often following the idea that charisma is given if someone could and would help others. Examples include influence (could help) and affability (would help) in scientific studies or power (could help), presence, and warmth (both would help) as a popular concept. Modelling high levels in these dimensions for humanoid robots or virtual agents, seems accomplishable. Beyond, also automatic measurement appears quite feasible with the recent advances in the related fields of Affective Computing and Social Signal Processing. Here, we, thereforem present a blueprint for building machines that can appear charismatic, but also analyse the charisma of others. To this end, we first provide the psychological perspective including different models of charisma and behavioural cues of it. We then switch to conversational charisma in spoken language as an exemplary modality that is essential for human-human and human-computer conversations. The computational perspective then deals with the recognition and generation of charismatic behaviour by AI. This includes an overview of the state of play in the field and the aforementioned blueprint. We then name exemplary use cases of computational charismatic skills before switching to ethical aspects and concluding this overview and perspective on building charisma-enabled AI.

Andreas Triantafyllopoulos, Sandra Zänkert, Alice Baird et al.

Stress is a major threat to well-being that manifests in a variety of physiological and mental symptoms. Utilising speech samples collected while the subject is undergoing an induced stress episode has recently shown promising results for the automatic characterisation of individual stress responses. In this work, we introduce new findings that shed light onto whether speech signals are suited to model physiological biomarkers, as obtained via cortisol measurements, or self-assessed appraisal and affect measurements. Our results show that different indicators impact acoustic features in a diverse way, but that their complimentary information can nevertheless be effectively harnessed by a multi-tasking architecture to improve prediction performance for all of them.

Liyizhe Peng, Zixing Zhang, Tao Pang et al.

The advent of large language models (LLMs) has gained tremendous attention over the past year. Previous studies have shown the astonishing performance of LLMs not only in other tasks but also in emotion recognition in terms of accuracy, universality, explanation, robustness, few/zero-shot learning, and others. Leveraging the capability of LLMs inevitably becomes an essential solution for emotion recognition. To this end, we further comprehensively investigate how LLMs perform in linguistic emotion recognition if we concentrate on this specific task. Specifically, we exemplify a publicly available and widely used LLM -- Chat General Language Model, and customise it for our target by using two different modal adaptation techniques, i.e., deep prompt tuning and low-rank adaptation. The experimental results obtained on six widely used datasets present that the adapted LLM can easily outperform other state-of-the-art but specialised deep models. This indicates the strong transferability and feasibility of LLMs in the field of emotion recognition.

Yi Chang, Sofiane Laridi, Zhao Ren et al.

Due to the development of machine learning and speech processing, speech emotion recognition has been a popular research topic in recent years. However, the speech data cannot be protected when it is uploaded and processed on servers in the internet-of-things applications of speech emotion recognition. Furthermore, deep neural networks have proven to be vulnerable to human-indistinguishable adversarial perturbations. The adversarial attacks generated from the perturbations may result in deep neural networks wrongly predicting the emotional states. We propose a novel federated adversarial learning framework for protecting both data and deep neural networks. The proposed framework consists of i) federated learning for data privacy, and ii) adversarial training at the training stage and randomisation at the testing stage for model robustness. The experiments show that our proposed framework can effectively protect the speech data locally and improve the model robustness against a series of adversarial attacks.

Andreas Triantafyllopoulos, Sandra Ottl, Alexander Gebhard et al.

Although running is a common leisure activity and a core training regiment for several athletes, between $29\%$ and $79\%$ of runners sustain an overuse injury each year. These injuries are linked to excessive fatigue, which alters how someone runs. In this work, we explore the feasibility of modelling the Borg received perception of exertion (RPE) scale (range: $[6-20]$), a well-validated subjective measure of fatigue, using audio data captured in realistic outdoor environments via smartphones attached to the runners' arms. Using convolutional neural networks (CNNs) on log-Mel spectrograms, we obtain a mean absolute error of $2.35$ in subject-dependent experiments, demonstrating that audio can be effectively used to model fatigue, while being more easily and non-invasively acquired than by signals from other sensors.

Xiangheng He, Andreas Triantafyllopoulos, Alexander Kathan et al.

Previous studies have shown the correlation between sensor data collected from mobile phones and human depression states. Compared to the traditional self-assessment questionnaires, the passive data collected from mobile phones is easier to access and less time-consuming. In particular, passive mobile phone data can be collected on a flexible time interval, thus detecting moment-by-moment psychological changes and helping achieve earlier interventions. Moreover, while previous studies mainly focused on depression diagnosis using mobile phone data, depression forecasting has not received sufficient attention. In this work, we extract four types of passive features from mobile phone data, including phone call, phone usage, user activity, and GPS features. We implement a long short-term memory (LSTM) network in a subject-independent 10-fold cross-validation setup to model both a diagnostic and a forecasting tasks. Experimental results show that the forecasting task achieves comparable results with the diagnostic task, which indicates the possibility of forecasting depression from mobile phone sensor data. Our model achieves an accuracy of 77.0 % for major depression forecasting (binary), an accuracy of 53.7 % for depression severity forecasting (5 classes), and a best RMSE score of 4.094 (PHQ-9, range from 0 to 27).

Davide Pigoli, Kieran Baker, Jobie Budd et al.

Since early in the coronavirus disease 2019 (COVID-19) pandemic, there has been interest in using artificial intelligence methods to predict COVID-19 infection status based on vocal audio signals, for example cough recordings. However, existing studies have limitations in terms of data collection and of the assessment of the performances of the proposed predictive models. This paper rigorously assesses state-of-the-art machine learning techniques used to predict COVID-19 infection status based on vocal audio signals, using a dataset collected by the UK Health Security Agency. This dataset includes acoustic recordings and extensive study participant meta-data. We provide guidelines on testing the performance of methods to classify COVID-19 infection status based on acoustic features and we discuss how these can be extended more generally to the development and assessment of predictive methods based on public health datasets.

Andreas Triantafyllopoulos, Meishu Song, Zijiang Yang et al.

In this work, we explore a novel few-shot personalisation architecture for emotional vocalisation prediction. The core contribution is an `enrolment' encoder which utilises two unlabelled samples of the target speaker to adjust the output of the emotion encoder; the adjustment is based on dot-product attention, thus effectively functioning as a form of `soft' feature selection. The emotion and enrolment encoders are based on two standard audio architectures: CNN14 and CNN10. The two encoders are further guided to forget or learn auxiliary emotion and/or speaker information. Our best approach achieves a CCC of $.650$ on the ExVo Few-Shot dev set, a $2.5\%$ increase over our baseline CNN14 CCC of $.634$.

Alexander Kathan, Andreas Triantafyllopoulos, Xiangheng He et al.

Digital health applications are becoming increasingly important for assessing and monitoring the wellbeing of people suffering from mental health conditions like depression. A common target of said applications is to predict the results of self-assessed Patient-Health-Questionnaires (PHQ), indicating current symptom severity of depressive individuals. In this work, we explore the potential of using actively-collected data to predict and forecast daily PHQ-2 scores on a newly-collected longitudinal dataset. We obtain a best MAE of 1.417 for daily prediction of PHQ-2 scores, which specifically in the used dataset have a range of 0 to 12, using leave-one-subject-out cross-validation, as well as a best MAE of 1.914 for forecasting PHQ-2 scores using data from up to the last 7 days. This illustrates the additive value that can be obtained by incorporating actively-collected data in a depression monitoring application.

Manuel Milling, Shuo Liu, Andreas Triantafyllopoulos et al.

Neural network models for audio tasks, such as automatic speech recognition (ASR) and acoustic scene classification (ASC), are susceptible to noise contamination for real-life applications. To improve audio quality, an enhancement module, which can be developed independently, is explicitly used at the front-end of the target audio applications. In this paper, we present an end-to-end learning solution to jointly optimise the models for audio enhancement (AE) and the subsequent applications. To guide the optimisation of the AE module towards a target application, and especially to overcome difficult samples, we make use of the sample-wise performance measure as an indication of sample importance. In experiments, we consider four representative applications to evaluate our training paradigm, i.e., ASR, speech command recognition (SCR), speech emotion recognition (SER), and ASC. These applications are associated with speech and non-speech tasks concerning semantic and non-semantic features, transient and global information, and the experimental results indicate that our proposed approach can considerably boost the noise robustness of the models, especially at low signal-to-noise ratios (SNRs), for a wide range of computer audition tasks in everyday-life noisy environments.

Vincent Karas, Andreas Triantafyllopoulos, Meishu Song et al.

Vocal bursts play an important role in communicating affect, making them valuable for improving speech emotion recognition. Here, we present our approach for classifying vocal bursts and predicting their emotional significance in the ACII Affective Vocal Burst Workshop & Challenge 2022 (A-VB). We use a large self-supervised audio model as shared feature extractor and compare multiple architectures built on classifier chains and attention networks, combined with uncertainty loss weighting strategies. Our approach surpasses the challenge baseline by a wide margin on all four tasks.

Björn W. Schuller, Dagmar M. Schuller

Emotion and a broader range of affective driver states can be a life decisive factor on the road. While this aspect has been investigated repeatedly, the advent of autonomous automobiles puts a new perspective on the role of computer-based emotion recognition in the car -- the passenger's one. This includes amongst others the monitoring of wellbeing during the commute such as to adjust the driving style or to adapt the info- and entertainment. This contribution aims to foresee according challenges and provide potential avenues towards affect modelling in a multimodal "audiovisual plus x" on the road context. From the technical end, this concerns holistic passenger modelling and reliable diarisation of the individuals in a vehicle. In conclusion, automated affect analysis has just matured to the point of applicability in autonomous vehicles in first selected use-cases, which will be discussed towards the end.

Andreas Triantafyllopoulos, Anastasia Semertzidou, Meishu Song et al.

More than two years after its outbreak, the COVID-19 pandemic continues to plague medical systems around the world, putting a strain on scarce resources, and claiming human lives. From the very beginning, various AI-based COVID-19 detection and monitoring tools have been pursued in an attempt to stem the tide of infections through timely diagnosis. In particular, computer audition has been suggested as a non-invasive, cost-efficient, and eco-friendly alternative for detecting COVID-19 infections through vocal sounds. However, like all AI methods, also computer audition is heavily dependent on the quantity and quality of available data, and large-scale COVID-19 sound datasets are difficult to acquire -- amongst other reasons -- due to the sensitive nature of such data. To that end, we introduce the COVYT dataset -- a novel COVID-19 dataset collected from public sources containing more than 8 hours of speech from 65 speakers. As compared to other existing COVID-19 sound datasets, the unique feature of the COVYT dataset is that it comprises both COVID-19 positive and negative samples from all 65 speakers. We analyse the acoustic manifestation of COVID-19 on the basis of these perfectly speaker characteristic balanced `in-the-wild' data using interpretable audio descriptors, and investigate several classification scenarios that shed light into proper partitioning strategies for a fair speech-based COVID-19 detection.

Yupei Li, Qiyang Sun, Hanqian Li et al.

Machine-generated music (MGM) has become a groundbreaking innovation with wide-ranging applications, such as music therapy, personalised editing, and creative inspiration within the music industry. However, the unregulated proliferation of MGM presents considerable challenges to the entertainment, education, and arts sectors by potentially undermining the value of high-quality human compositions. Consequently, MGM detection (MGMD) is crucial for preserving the integrity of these fields. Despite its significance, MGMD domain lacks comprehensive systematic evaluation results necessary to drive meaningful progress. To address this gap, we conduct experiments on existing large-scale datasets using a range of foundational models for audio processing, establishing systematic evaluation results tailored to the MGMD task. Our selection includes traditional machine learning models, deep neural networks, Transformer-based architectures, and State space models (SSM). Recognising the inherently multimodal nature of music, which integrates both melody and lyrics, we also explore fundamental multimodal models in our experiments. Beyond providing basic binary classification outcomes, we delve deeper into model behaviour using multiple explainable Artificial Intelligence (XAI) tools, offering insights into their decision-making processes. Our analysis reveals that ResNet18 performs the best according to in-domain and out-of-domain tests. By providing a comprehensive comparison of systematic evaluation results and their interpretability, we propose several directions to inspire future research to develop more robust and effective detection methods for MGM. We provide our codes and some samples on Github repository.

Zheng Lian, Haiyang Sun, Licai Sun et al.

Multimodal emotion recognition is an important research topic in artificial intelligence. Over the past few decades, researchers have made remarkable progress by increasing the dataset size and building more effective algorithms. However, due to problems such as complex environments and inaccurate annotations, current systems are hard to meet the demands of practical applications. Therefore, we organize the MER series of competitions to promote the development of this field. Last year, we launched MER2023, focusing on three interesting topics: multi-label learning, noise robustness, and semi-supervised learning. In this year's MER2024, besides expanding the dataset size, we further introduce a new track around open-vocabulary emotion recognition. The main purpose of this track is that existing datasets usually fix the label space and use majority voting to enhance the annotator consistency. However, this process may lead to inaccurate annotations, such as ignoring non-majority or non-candidate labels. In this track, we encourage participants to generate any number of labels in any category, aiming to describe emotional states as accurately as possible. Our baseline code relies on MERTools and is available at: https://github.com/zeroQiaoba/MERTools/tree/master/MER2024.

Alexander Gebhard, Andreas Triantafyllopoulos, Dominik Arend et al.

A soundscape is composed of three types of sound: biophony (sounds made by animals), geophony (natural abiotic sounds) and anthropophony (sounds made by humans). A key research question in the field of soundscape ecology is how these components interact with each other, specifically how biophony responds to geophony and anthropophony. Nevertheless, as of today, there are not many analytical instruments that enable the distinct quantification of these elements. Recent machine learning (ML) approaches aim to support automated analysis but often rely on task-specific or clean data, limiting generalisation to noisy passive acoustic monitoring (PAM) recordings. This study presents a clear and reproducible structure to build ML models for coarse soundscape classification and introduces CoarseSoundNet, a deep learning model trained to distinguish biophony, geophony, and anthropophony under realistic PAM conditions. We systematically investigate model architectures, the influence of an additional training class, data composition, and evaluation strategies. Our findings suggest that model performance improves with additional PAM data, especially when similar to the target domain, and by introducing an explicit silence class during training. Class-specific decision thresholds and duration-based constraints further enhance performance, particularly for anthropophony and geophony. Error analyses exhibit challenges for anthropophony due to masking effects and confusions for silence and insect sounds for geophony and biophony. Finally, we conduct an ecological case study which shows that pre-filtering recordings with CoarseSoundNet yields acoustic index trends comparable to ground-truth filtering, supporting its use as an effective preprocessing tool for ecoacoustic analyses.

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.

Andreas Triantafyllopoulos, Jakub Šťastný, Alexios Terpinas et al.

Reinforcement learning is a powerful learning paradigm that has spearheaded progress in numerous domains. Its core promise lies in learning through high-level goals without the need for granular labels. However, it still remains elusive in the realm of audio, where it has received substantially less attention than in computer vision or other domains. The key question remains: how can agents learn to listen purely via reward-driven exploration? In this contribution, we present an overview of previous attempts and a new conceptual framework for learning to listen by reward. Our approach depends on the continuous search for novel sound sources. We formulate our framework, discuss open technical challenges, and present a first proof-of-concept implementation that showcases the feasibility of our approach.

Yupei Li, Qiyang Sun, Sunil Munthumoduku Krishna Murthy et al.

Affective Computing (AC) is essential for advancing Artificial General Intelligence (AGI), with emotion recognition serving as a key component. However, human emotions are inherently dynamic, influenced not only by an individual's expressions but also by interactions with others, and single-modality approaches often fail to capture their full dynamics. Multimodal Emotion Recognition (MER) leverages multiple signals but traditionally relies on utterance-level analysis, overlooking the dynamic nature of emotions in conversations. Emotion Recognition in Conversation (ERC) addresses this limitation, yet existing methods struggle to align multimodal features and explain why emotions evolve within dialogues. To bridge this gap, we propose GatedxLSTM, a novel speech-text multimodal ERC model that explicitly considers voice and transcripts of both the speaker and their conversational partner(s) to identify the most influential sentences driving emotional shifts. By integrating Contrastive Language-Audio Pretraining (CLAP) for improved cross-modal alignment and employing a gating mechanism to emphasise emotionally impactful utterances, GatedxLSTM enhances both interpretability and performance. Additionally, the Dialogical Emotion Decoder (DED) refines emotion predictions by modelling contextual dependencies. Experiments on the IEMOCAP dataset demonstrate that GatedxLSTM achieves state-of-the-art (SOTA) performance among open-source methods in four-class emotion classification. These results validate its effectiveness for ERC applications and provide an interpretability analysis from a psychological perspective.

Chia-Hsin Lin, Charles Jones, Björn W. Schuller et al.

Despite significant advancements in deep learning for vision and natural language, unsupervised domain adaptation in audio remains relatively unexplored. We, in part, attribute this to the lack of an appropriate benchmark dataset. To address this gap, we present Synthia's melody, a novel audio data generation framework capable of simulating an infinite variety of 4-second melodies with user-specified confounding structures characterised by musical keys, timbre, and loudness. Unlike existing datasets collected under observational settings, Synthia's melody is free of unobserved biases, ensuring the reproducibility and comparability of experiments. To showcase its utility, we generate two types of distribution shifts-domain shift and sample selection bias-and evaluate the performance of acoustic deep learning models under these shifts. Our evaluations reveal that Synthia's melody provides a robust testbed for examining the susceptibility of these models to varying levels of distribution shift.

Jiaming Cheng, Ruiyu Liang, Ye Ni et al.

In this paper, we propose an intra-set and inter-set recursive fusion framework with time-frequency calibrated knowledge distillation (I$^2$SRF-TFCKD) for SE. Different from previous distillation strategies for SE, the proposed framework fully exploits the time-frequency differential information of speech while facilitating both local information focusing and global knowledge circulation. Firstly, we construct a collaborative distillation paradigm for intra-set and inter-set correlations. Within a correlated set, multi-layer teacher-student features are pairwise matched for calibrated distillation. Subsequently, we generate representative features from each correlated set through recursive fusion to form the fused feature set that enables inter-set knowledge interaction. Secondly, we propose a multi-layer interactive distillation based on dual-stream time-frequency cross-calibration, which calculates the teacher-student similarity calibration weights in the time and frequency domains respectively and performs cross-weighting, thus enabling refined allocation of distillation contributions across different layers according to speech characteristics. The proposed distillation strategy is applied to the dual-path dilated convolutional recurrent network (DPDCRN) that ranked first in the SE track of the L3DAS23 challenge. To evaluate the effectiveness of I$^2$SRF-TFCKD, we conduct experiments on both single-channel and multi-channel SE datasets. Objective evaluations demonstrate that the proposed KD strategy consistently and effectively improves the performance of the low-complexity student model and outperforms other distillation schemes.

Yupei Li, Shuaijie Shao, Manuel Milling et al.

Depression is a growing concern gaining attention in both public discourse and AI research. While deep neural networks (DNNs) have been used for recognition, they still lack real-world effectiveness. Large language models (LLMs) show strong potential but require domain-specific fine-tuning and struggle with non-textual cues. Since depression is often expressed through vocal tone and behaviour rather than explicit text, relying on language alone is insufficient. Diagnostic accuracy also suffers without incorporating psychological expertise. To address these limitations, we present, to the best of our knowledge, the first application of LLMs to multimodal depression detection using the DAIC-WOZ dataset. We extract the audio features using the pre-trained model Wav2Vec, and mapped it to text-based LLMs for further processing. We also propose a novel strategy for incorporating psychological knowledge into LLMs to enhance diagnostic performance, specifically using a question and answer set to grant authorised knowledge to LLMs. Our approach yields a notable improvement in both Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) compared to a base score proposed by the related original paper. The codes are available at https://github.com/myxp-lyp/Depression-detection.git

Simon Rampp, Manuel Milling, Andreas Triantafyllopoulos et al.

Curriculum learning (CL) describes a machine learning training strategy in which samples are gradually introduced into the training process based on their difficulty. Despite a partially contradictory body of evidence in the literature, CL finds popularity in deep learning research due to its promise of leveraging human-inspired curricula to achieve higher model performance. Yet, the subjectivity and biases that follow any necessary definition of difficulty, especially for those found in orderings derived from models or training statistics, have rarely been investigated. To shed more light on the underlying unanswered questions, we conduct an extensive study on the robustness and similarity of the most common scoring functions for sample difficulty estimation, as well as their potential benefits in CL, using the popular benchmark dataset CIFAR-10 and the acoustic scene classification task from the DCASE2020 challenge as representatives of computer vision and computer audition, respectively. We report a strong dependence of scoring functions on the training setting, including randomness, which can partly be mitigated through ensemble scoring. While we do not find a general advantage of CL over uniform sampling, we observe that the ordering in which data is presented for CL-based training plays an important role in model performance. Furthermore, we find that the robustness of scoring functions across random seeds positively correlates with CL performance. Finally, we uncover that models trained with different CL strategies complement each other by boosting predictive power through late fusion, likely due to differences in the learnt concepts. Alongside our findings, we release the aucurriculum toolkit (https://github.com/autrainer/aucurriculum), implementing sample difficulty and CL-based training in a modular fashion.