Learning Paralinguistic Features from Audiobooks through Style Voice Conversion

Abstract

Paralinguistics, the non-lexical components of speech, play a crucial role in human-human interaction. Models designed to recognize paralinguistic information, particularly speech emotion and style, are difficult to train because of the limited labeled datasets available. In this work, we present a new framework that enables a neural network to learn to extract paralinguistic attributes from speech using data that are not annotated for emotion. We assess the utility of the learned embeddings on the downstream tasks of emotion recognition and speaking style detection, demonstrating significant improvements over surface acoustic features as well as over embeddings extracted from other unsupervised approaches. Our work enables future systems to leverage the learned embedding extractor as a separate component capable of highlighting the paralinguistic components of speech.

Evaluating Self-Supervised Speech Representations for Speech Emotion Recognition

“…Aldeneh et al [11] proposed a framework to learn to extract paralinguistic embedding. The authors showed that converting synthetic-neutral speech to expressive speech based on that embedding improved the results from acoustic features and other evaluated embeddings.…”