Generative Approach Using Soft-Labels to Learn Uncertainty in Predicting Emotional Attributes

Label Uncertainty Modeling and Prediction for Speech Emotion Recognition using t-Distributions

“…sciences and SER has shown increasing prominence in highcritical and socially relevant domains, e.g. health, security, and employee well-being [2], [4], [5].…”

“…In SER, label uncertainty has been approached using softlabels [5], multi-task learning (MTL) [9], [10], stochastic models [11], [12], and label distribution learning [13], [14]. The subjective annotations of emotion creates a label distribution which explains the uncertainty in emotions [5].…”

“…In SER, label uncertainty has been approached using softlabels [5], multi-task learning (MTL) [9], [10], stochastic models [11], [12], and label distribution learning [13], [14]. The subjective annotations of emotion creates a label distribution which explains the uncertainty in emotions [5]. In this light, label distribution learning techniques for label uncertainty in SER are gaining research focus, with improved performances [13], [14].…”

“…In this light, label distribution learning techniques for label uncertainty in SER are gaining research focus, with improved performances [13], [14]. However, a problem with label distribution learning in SER is the limited annotations available [5], [14], due to resource inefficient task of gaining more annotations [5].…”

“…However, a Gaussian assumption with limited samples is not well justified [15], [16], as the central limit theorem (CLT) which primarily backs Gaussian distributions does not hold with insufficient numbers of sam-ples [17]. Publicly available SER datasets commonly comprise of only three to six annotations [18]- [22], and well agree that gaining more annotations is resource inefficient [5], [23]. In this light, it is important for machine learning (ML) models to account for the limited annotations and model the label distribution accordingly.…”