The lack of a large, natural emotional database is one of the key barriers to translate results on speech emotion recognition in controlled conditions into real-life applications. Collecting emotional databases is expensive and time demanding, which limits the size of existing corpora. Current approaches used to collect spontaneous databases tend to provide unbalanced emotional content, which is dictated by the given recording protocol (e.g., positive for colloquial conversations, negative for discussion or debates). The size and speaker diversity are also limited. This paper proposes a novel approach to effectively build a large, naturalistic emotional database with balanced emotional content, reduced cost and reduced manual labor. It relies on existing spontaneous recordings obtained from audio-sharing websites. The proposed approach combines machine learning algorithms to retrieve recordings conveying balanced emotional content with a cost effective annotation process using crowdsourcing, which make it possible to build a large scale speech emotional database. This approach provides natural emotional renditions from multiple speakers, with different channel conditions and conveying balanced emotional content that are difficult to obtain with alternative data collection protocols.

The performance of speech emotion recognition is affected by the differences in data distributions between train (source domain) and test (target domain) sets used to build and evaluate the models. This is a common problem, as multiple studies have shown that the performance of emotional classifiers drops when they are exposed to data that do not match the distribution used to build the emotion classifiers. The difference in data distributions becomes very clear when the training and testing data come from different domains, causing a large performance gap between development and testing performance. Due to the high cost of annotating new data and the abundance of unlabeled data, it is crucial to extract as much useful information as possible from the available unlabeled data. This study looks into the use of adversarial multitask training to extract a common representation between train and test domains. The primary task is to predict emotional-attribute-based descriptors for arousal, valence, or dominance. The secondary task is to learn a common representation, where the train and test domains cannot be distinguished. By using a gradient reversal layer, the gradients coming from the domain classifier are used to bring the source and target domain representations closer. We show that exploiting unlabeled data consistently leads to better emotion recognition performance across all emotional dimensions. We visualize the effect of adversarial training on the feature representation across the proposed deep learning architecture. The analysis shows that the data representations for the train and test domains converge as the data are passed to deeper layers of the network. We also evaluate the difference in performance when we use a shallow neural network versus a deep neural network and the effect of the number of shared layers used by the task and domain classifiers.

Representing computationally everyday emotional states is a challenging task and, arguably, one of the most fundamental for affective computing. Standard practice in emotion annotation is to ask humans to assign a value of intensity or a class value to each emotional behavior they observe. Psychological theories and evidence from multiple disciplines including neuroscience, economics and artificial intelligence, however, suggest that the task of assigning reference-based values to subjective notions is better aligned with the underlying representations. This paper draws together the theoretical reasons to favor ordinal labels for representing and annotating emotion, reviewing the literature across several disciplines. We go on to discuss good and bad practices of treating ordinal and other forms of annotation data and make the case for preference learning methods as the appropriate approach for treating ordinal labels. We finally discuss the advantages of ordinal annotation with respect to both reliability and validity through a number of case studies in affective computing, and address common objections to the use of ordinal data. Overall, the thesis that emotions are by nature ordinal is supported by both theoretical arguments and evidence, and opens new horizons for the way emotions are viewed, represented and analyzed computationally.

Speech emotion recognition (SER) systems find applications in various fields such as healthcare, education, and security and defense. A major drawback of these systems is their lack of generalization across different conditions. For example, systems that show superior performance on certain databases show poor performance when tested on other corpora. This problem can be solved by training models on large amounts of labeled data from the target domain, which is expensive and time-consuming. Another approach is to increase the generalization of the models. An effective way to achieve this goal is by regularizing the models through multitask learning (MTL), where auxiliary tasks are learned along with the primary task. These methods often require the use of labeled data which is computationally expensive to collect for emotion recognition (gender, speaker identity, age or other emotional descriptors). This study proposes the use of ladder networks for emotion recognition, which utilizes an unsupervised auxiliary task. The primary task is a regression problem to predict emotional attributes. The auxiliary task is the reconstruction of intermediate feature representations using a denoising autoencoder. This auxiliary task does not require labels so it is possible to train the framework in a semi-supervised fashion with abundant unlabeled data from the target domain. This study shows that the proposed approach creates a powerful framework for SER, achieving superior performance than fully supervised single-task learning (STL) and MTL baselines. We implement the approach with sentence-level or frame-level features, demonstrating the flexibility of our approach. Additionally, the generalization of the ladder networks is evaluated in cross-corpus settings using sentence-level features, obtaining important improvements. Compared to the STL baselines, the proposed approach achieves relative gains in concordance correlation coefficient (CCC) between 3.0% and 3.5% for within corpus evaluations, and between 16.1% and 74.1% for cross corpus evaluations, highlighting the power of the architecture.

This study introduces a method to design a curriculum for machine-learning to maximize the efficiency during the training process of deep neural networks (DNNs) for speech emotion recognition. Previous studies in other machine-learning problems have shown the benefits of training a classifier following a curriculum where samples are gradually presented in increasing level of difficulty. For speech emotion recognition, the challenge is to establish a natural order of difficulty in the training set to create the curriculum. We address this problem by assuming that ambiguous samples for humans are also ambiguous for computers. Speech samples are often annotated by multiple evaluators to account for differences in emotion perception across individuals. While some sentences with clear emotional content are consistently annotated, sentences with more ambiguous emotional content present important disagreement between individual evaluations. We propose to use the disagreement between evaluators as a measure of difficulty for the classification task. We propose metrics that quantify the inter-evaluation agreement to define the curriculum for regression problems and binary and multi-class classification problems. The experimental results consistently show that relying on a curriculum based on agreement between human judgments leads to statistically significant improvements over baselines trained without a curriculum.

An automatic speech recognition (ASR) system is a key component in current speech-based systems. However, the surrounding acoustic noise can severely degrade the performance of an ASR system. An appealing solution to address this problem is to augment conventional audio-based ASR systems with visual features describing lip activity. This paper proposes a novel end-to-end, multitask learning (MTL), audiovisual ASR (AV-ASR) system. A key novelty of the approach is the use of MTL, where the primary task is AV-ASR, and the secondary task is audiovisual voice activity detection (AV-VAD). We obtain a robust and accurate audiovisual system that generalizes across conditions. By detecting segments with speech activity, the AV-ASR performance improves as its connectionist temporal classification (CTC) loss function can leverage from the AV-VAD alignment information. Furthermore, the end-to-end system learns from the raw audiovisual inputs a discriminative high-level representation for both speech tasks, providing the flexibility to mine information directly from the data. The proposed architecture considers the temporal dynamics within and across modalities, providing an appealing and practical fusion scheme. We evaluate the proposed approach on a large audiovisual corpus (over 60 hours), which contains different channel and environmental conditions, comparing the results with competitive single task learning (STL) and MTL baselines. Although our main goal is to improve the performance of our ASR task, the experimental results show that the proposed approach can achieve the best performance across all conditions for both speech tasks. In addition to state-of-the-art performance in AV-ASR, the proposed solution can also provide valuable information about speech activity, solving two of the most important tasks in speech-based applications.