A Comparative Study of Neural Models for Emotional Voice Conversion

Abstract

Emotional voice conversion (EVC) is a fundamental task in the field of speech processing, enabling the modification of the emotional content of a spoken message while preserving the speaker's identity. This article presents a preliminary comparative study of various neural models applied to emotional voice conversion. We explore the performance of state-of-the-art models: seq2seq-EVC and CycleGAN-EVC. To do so, we evaluate these models on various emotional datasets and analyze their ability to accurately convert emotions across a spectrum of affective states. Our preliminary experiments reveal insights into the strengths and limitations of each neural model in capturing and transferring emotional nuances in speech. We discuss key factors such as model architecture, dataset size, and training strategies, shedding light on the trade-offs between computational complexity and conversion quality. Additionally, we assess the perceptual quality of the converted emotional voice using subjective metrics, providing a complete view of model performance.