Addressing Variability in Speech when Recognizing Emotion and Mood In-the-Wild

Abstract

Bipolar disorder is a chronic mental illness, affecting 4% of Americans, that is characterized by periodic mood changes ranging from severe depression to extreme compulsive highs. Both mania and depression profoundly impact the behavior of affected individuals, resulting in potentially devastating personal and social consequences. Bipolar disorder is managed clinically with regular interactions with care providers, who assess mood, energy levels, and the form and content of speech. Recent work has proposed smartphones for automatically monitoring mood using speech.

Much of the early work in speech-centered mood detection has been done in the laboratory or clinic and is not reflective of the variability found in real-world conversations and conditions. Outside of these settings, automatic mood detection is hard, as the recordings include environmental noise, differences in recording devices, and variations in subject speaking patterns. Without addressing these issues, it is difficult to move towards a passive mobile health system. My research works to address this variability present in speech so that such a system can be created, allowing for interventions to mitigate the life-changing effects of mood transitions.

However detecting mood directly from speech is difficult, as mood varies over the course of days or weeks, while speech fluctuates rapidly. To address this, my thesis explores how an intermediate step can be used to aid in this prediction. For example, one of the major symptoms of bipolar disorder is emotion dysregulation - changes in the way emotions are perceived and a lack of inhibition in their expression. My work has supported the relationship between automatically extracted emotion estimates and mood. Because of this, my thesis explores how to mitigate the variability found when detecting emotion from speech. The remainder of my thesis is focused on employing these emotion-based features, as well as features based on language content, to real-world applications. This dissertation is divided into the following parts:

Part I: I address the direct classification of mood from speech. This is accomplished by addressing variability due to recording device using preprocessing and multi-task learning. I then show how both subject-specific and population-general information can be combined to significantly improve mood detection.

Part II: I explore the automatic detection of emotion from speech and how to control for the other factors of variability present in the speech signal. I use progressive networks as a method to augment emotion with other paralinguistic data including gender and speaker, as well as other datasets. Additionally, I introduce a novel domain generalization method for cross-corpus detection.

Part III: I demonstrate real-world applications of speech mood monitoring using everyday conversations. I show how the previously introduced generalized model can predict emotion from the speech of individuals with suicidal ideation, demonstrating its effectiveness across domains. Furthermore, I use these predictions to distinguish individuals with suicidal thoughts from healthy controls. Lastly, I introduce a novel framework for intervention detection in individuals with bipolar disorder. I then create a natural speech mood monitoring system based on features derived from measures of emotion and automatic speech recognition (ASR) transcripts and show effective intervention detection.

I conclude this dissertation with the following future directions: (1) Extending my emotion generalization system to include multiple modalities and factors of variability; (2) Expanding natural speech mood monitoring by including more devices, exploring other data besides speech, and investigating mood rating causality.