Self-supervised pretraining has transformed speech representation learning, enabling models to generalize across various downstream tasks. However, empirical studies have highlighted two notable gaps. First, different speech tasks require varying levels of acoustic and semantic information, which are encoded at different layers within the model. This adds the extra complexity of layer selection on downstream tasks to reach optimal performance. Second, the entanglement of acoustic and semantic information can undermine model robustness, particularly in varied acoustic environments. To address these issues, we propose a two-branch multitask finetuning strategy that integrates Automatic Speech Recognition and transcript-aligned audio reconstruction, designed to preserve and disentangle semantic and acoustic information in a final layer of a pretrained model. Experiments with the pretrained Wav2Vec 2.0 model demonstrate that our approach surpasses ASR-only finetuning across multiple downstream tasks, and it significantly improves ASR robustness in acoustically varied (emotional) speech.

Modern affective computing systems rely heavily on datasets with human-annotated emotion labels for both training and evaluation. However, human annotations are expensive to obtain, sensitive to study design, and difficult to quality control, because of the subjective nature of emotions. Meanwhile, Large Language Models (LLMs) have shown remarkable performance on many Natural Language Understanding tasks, emerging as a promising tool for text annotation. In this work, we analyze the complexities of emotion annotation in the context of LLMs, focusing on GPT-4 as a leading model. In our experiments, GPT-4 achieves high ratings in a human evaluation study, painting a more positive picture than previous work, in which human labels served as the only ground truth. On the other hand, we observe differences between human and GPT-4 emotion perception, underscoring the importance of human input in annotation studies. To harness GPT-4's strength while preserving human perspective, we explore two ways of integrating GPT-4 into emotion annotation pipelines, showing its potential to flag low-quality labels, reduce the workload of human annotators, and improve downstream model learning performance and efficiency. Together, our findings highlight opportunities for new emotion labeling practices and suggest the use of LLMs as a promising tool to aid human annotation.

Humans’ internal states play a key role in human-machine interaction, leading to the rise of human state estimation as a prominent field. Compared to swift state changes such as surprise and irritation, modeling gradual states like trust and satisfaction are further challenged by label sparsity: long time-series signals are usually associated with a single label, making it difficult to identify the critical span of state shifts. Windowing has been one widely-used technique to enable localized analysis of long time-series data. However, the performance of downstream models can be sensitive to the window size, and determining the optimal window size demands domain expertise and extensive search. To address this challenge, we propose a Selective Windowing Attention Network (SWAN), which employs window prompts and masked attention transformation to enable the selection of attended intervals with flexible lengths. We evaluate SWAN on the task of trust prediction on a new multimodal driving simulation dataset. Experiments show that SWAN significantly outperforms an existing empirical window selection baseline and neural network baselines including CNN-LSTM and Transformer. Furthermore, it shows robustness across a wide span of windowing ranges, compared to the traditional windowing approach.

A speech disfluency, such as a filled pause, repetition, or revision, disrupts the typical flow of speech. Disfluency modeling has grown as a research area, as recent work has shown that these disfluencies may help in assessing health conditions. For example, for individuals with cognitive impairment, changes in disfluencies may indicate worsening symptoms. However, work on disfluency modeling has focused heavily on detection and less on categorization. Work that has focused on categorization has suffered with two specific classes: repetitions and revisions. In this paper, we evaluate how BERT (Bidirectional Encoder Representations from Transformers) compares to other models on disfluency detection and categorization. We also propose adding a second fine-tuning task where BERT learns to distance repetitions and revisions from their repairs with triplet loss. We find that BERT and BERT with triplet loss outperform previous work on disfluency detection and categorization, particularly for repetitions and revisions. In this paper we present the first analysis of how these models can be fine-tuned on widely available disfluency data, and then used in an off-the-shelf manner on small corpora of pathological speech.