Publications

The dissertation studies affect, a central social signal, from two sides. On the recognition side it contributes probabilistic methods that model label uncertainty, combining Bayesian neural networks with label distribution learning and replacing Gaussian assumptions with Student's t-distributions when annotations are sparse; a protocol for annotating group-level affect over time; and graph-based multimodal models of collective affect. On the synthesis side it proposes generative models for emotion conversion in in-the-wild speech without parallel data, explores diffusion decoders for controllable affect, and shows the value of explicit duration modeling. Across both, ground truth is treated as a distribution rather than an absolute.

A summary. The full abstract and PDF are on eDiss.

Collaborating in a group, whether face-to-face or virtually, involves continuously expressing emotions and interpreting those of other group members. Therefore, understanding group affect is essential to comprehending how groups interact and succeed in collaborative efforts. In this study, we move beyond individual-level affect and investigate group-level affect — a collective phenomenon that reflects the shared mood or emotions among group members at a particular moment. As the first in literature, we gather annotations for group-level affective expressions using a fine-grained temporal approach (15 second windows) that also captures the inherent dynamics of the collective construct. To this end, we use trained annotators and an annotation procedure specifically tuned to capture the entire scope of the group interaction. In addition, we model group affect dynamics over time. One way to study the ebb and flow of group affect in group interactions is to model the underlying convergence (driven by emotional contagion) and divergence (resulting from emotional reactivity) of affective expressions amongst group members. To capture these interpersonal dynamics, we extract synchrony based features from both audio and visual social signal cues. An analysis of these features reveals that interacting groups tend to diverge in terms of their social signals along neutral levels of group affect, and converge along extreme levels of affect expression. We further present results on the predictive modeling of dynamic group affect which underscores the importance of using synchrony-based features in the modeling process, as well as the multimodal nature of group affect. We anticipate that the presented models will serve as the baselines of future research on the automatic recognition of dynamic group affect.

Abstract of the arXiv preprint.

Speech Emotion Conversion aims to modify the emotion expressed in input speech while preserving lexical content and speaker identity. Recently, generative modeling approaches have shown promising results in changing local acoustic properties such as fundamental frequency, spectral envelope and energy, but often lack the ability to control the duration of sounds. To address this, we propose a duration modeling framework using resynthesis-based discrete content representations, enabling modification of speech duration to reflect target emotions and achieve controllable speech rates without using parallel data. Experimental results reveal that the inclusion of the proposed duration modeling framework significantly enhances emotional expressiveness, in the in-the-wild MSP-Podcast dataset. Analyses show that low-arousal emotions correlate with longer durations and slower speech rates, while high-arousal emotions produce shorter, faster speech.

Speech emotion conversion is the task of converting the expressed emotion of a spoken utterance to a target emotion while preserving the lexical content and speaker identity. While most existing works in speech emotion conversion rely on acted-out datasets and parallel data samples, in this work we specifically focus on more challenging in-the-wild scenarios and do not rely on parallel data. To this end, we propose a diffusion-based generative model for speech emotion conversion, the EmoConv-Diff, that is trained to reconstruct an input utterance while also conditioning on its emotion. Subsequently, at inference, a target emotion embedding is employed to convert the emotion of the input utterance to the given target emotion. As opposed to performing emotion conversion on categorical representations, we use a continuous arousal dimension to represent emotions while also achieving intensity control. We validate the proposed methodology on a large in-the-wild dataset, the MSP-Podcast v1.10. Our results show that the proposed diffusion model is indeed capable of synthesizing speech with a controllable target emotion. Crucially, the proposed approach shows improved performance along the extreme values of arousal and thereby addresses a common challenge in the speech emotion conversion literature.

To train machine learning algorithms to predict emotional expressions in terms of arousal and valence, annotated datasets are needed. However, as different people perceive others' emotional expressions differently, their annotations are subjective. To account for this, annotations are typically collected from multiple annotators and averaged to obtain ground-truth labels. However, when exclusively trained on this averaged ground-truth, the model is agnostic to the inherent subjectivity in emotional expressions. In this work, we therefore propose an end-to-end Bayesian neural network capable of being trained on a distribution of annotations to also capture the subjectivity-based label uncertainty. Instead of a Gaussian, we model the annotation distribution using Student's t-distribution, which also accounts for the number of annotations available. We derive the corresponding Kullback-Leibler divergence loss and use it to train an estimator for the annotation distribution, from which the mean and uncertainty can be inferred. We validate the proposed method using two in-the-wild datasets. We show that the proposed t-distribution based approach achieves state-of-the-art uncertainty modeling results in speech emotion recognition, and also consistent results in cross-corpora evaluations. Furthermore, analyses reveal that the advantage of a t-distribution over a Gaussian grows with increasing inter-annotator correlation and a decreasing number of annotations available.

The quality of daily spontaneous conversations is of importance towards both our well-being as well as the development of interactive social agents. Prior research directly studying the quality of social conversations has operationalized it in narrow terms, associating greater quality to less small talk. Other works taking a broader perspective of interaction experience have indirectly studied quality through one of the several overlapping constructs such as rapport or engagement, in isolation. In this work we bridge this gap by proposing a holistic conceptualization of conversation quality, building upon the collaborative attributes of cooperative conversation floors. Taking a multilevel perspective of conversation, we develop and validate two instruments for perceived conversation quality (PCQ) at the individual and group levels. Specifically, we motivate capturing external raters' gestalt impressions of participant experiences from thin slices of behavior, and collect annotations of PCQ on the publicly available MatchNMingle dataset of in-the-wild mingling conversations. Finally, we present an analysis of behavioral features that are predictive of PCQ. We find that for the conversations in MatchNMingle, raters tend to associate smaller group sizes, equitable speaking turns with fewer interruptions, and time taken for synchronous bodily coordination with higher PCQ.

Speech emotion conversion aims to convert the expressed emotion of a spoken utterance to a target emotion while preserving the lexical information and the speaker's identity. In this work, we specifically focus on in-the-wild emotion conversion where parallel data does not exist, and the problem of disentangling lexical, speaker, and emotion information arises. In this paper, we introduce a methodology that uses self-supervised networks to disentangle the lexical, speaker, and emotional content of the utterance, and subsequently uses a HiFiGAN vocoder to resynthesise the disentangled representations to a speech signal of the targeted emotion. For better representation and to achieve emotion intensity control, we specifically focus on the arousal dimension of continuous representations, as opposed to performing emotion conversion on categorical representations. We test our methodology on the large in-the-wild MSP-Podcast dataset. Results reveal that the proposed approach is aptly conditioned on the emotional content of input speech and is capable of synthesising natural-sounding speech for a target emotion. Results further reveal that the methodology better synthesises speech for mid-scale arousal (2 to 6) than for extreme arousal (1 and 7).

In large part due to their implicit semantic modeling, self-supervised learning (SSL) methods have significantly increased the performance of valence recognition in speech emotion recognition (SER) systems. Yet, their large size may often hinder practical implementations. In this work, we take HuBERT as an example of an SSL model and analyze the relevance of each of its layers for SER. We show that shallow layers are more important for arousal recognition while deeper layers are more important for valence. This observation motivates the importance of additional textual information for accurate valence recognition, as the distilled framework lacks the depth of its large-scale SSL teacher. Thus, we propose an audio-textual distilled SSL framework that, while having only ~20% of the trainable parameters of a large SSL model, achieves on par performance across the three emotion dimensions (arousal, valence, dominance) on the MSP-Podcast v1.10 dataset.

As different people perceive others' emotional expressions differently, their annotation in terms of arousal and valence are per se subjective. To address this, these emotion annotations are typically collected by multiple annotators and averaged across annotators in order to obtain labels for arousal and valence. However, besides the average, also the uncertainty of a label is of interest, and should also be modeled and predicted for automatic emotion recognition. In the literature, for simplicity, label uncertainty modeling is commonly approached with a Gaussian assumption on the collected annotations. However, as the number of annotators is typically rather small due to resource constraints, we argue that the Gaussian approach is a rather crude assumption. In contrast, in this work we propose to model the label distribution using a Student's t-distribution which allows us to account for the number of annotations available. With this model, we derive the corresponding Kullback-Leibler divergence based loss function and use it to train an estimator for the distribution of emotion labels, from which the mean and uncertainty can be inferred. Through qualitative and quantitative analysis, we show the benefits of the t-distribution over a Gaussian distribution. Results reveal that our t-distribution based approach improves over the Gaussian approach with state-of-the-art uncertainty modeling results in speech-based emotion recognition, along with an optimal and even faster convergence.

Emotions are subjective constructs. Recent end-to-end speech emotion recognition systems are typically agnostic to the subjective nature of emotions, despite their state-of-the-art performance. In this work, we introduce an end-to-end Bayesian neural network architecture to capture the inherent subjectivity in the arousal dimension of emotional expressions. To the best of our knowledge, this work is the first to use Bayesian neural networks for speech emotion recognition. At training, the network learns a distribution of weights to capture the inherent uncertainty related to subjective arousal annotations. To this end, we introduce a loss term that enables the model to be explicitly trained on a distribution of annotations, rather than training them exclusively on mean or gold-standard labels. We evaluate the proposed approach on the AVEC'16 dataset. Qualitative and quantitative analysis of the results reveals that the proposed model can aptly capture the distribution of subjective arousal annotations, with state-of-the-art results in mean and standard deviation estimations for uncertainty modeling.

Social interactions in general are multifaceted and there exists a wide set of factors and events that influence them. In this paper, we quantify social interactions with a holistic viewpoint on individual experiences, particularly focusing on non-task-directed spontaneous interactions. To achieve this, we design a novel perceived measure, the perceived Conversation Quality, which intends to quantify spontaneous interactions by accounting for several socio-dimensional aspects of individual experiences. To further quantitatively study spontaneous interactions, we devise a questionnaire which measures the perceived Conversation Quality, at both the individual- and at the group- level. Using the questionnaire, we collected perceived annotations for conversation quality in a publicly available dataset using naive annotators. The results of the analysis performed on the distribution and the inter-annotator agreeability shows that naive annotators tend to agree less in cases of low conversation quality samples, especially while annotating for group-level conversation quality.