An Overview of Affective Speech Synthesis and Conversion in the Deep Learning Era

End-To-End Label Uncertainty Modeling for Speech-based Arousal Recognition Using Bayesian Neural Networks

“…More importantly, to improve the naturalness of machine communication, the generation of emotionally expressive speech is required. While speech generation technologies have been making significant progress [4,5], emotionally expressive speech generation is still a challenge [6,7]. Speech emotion conversion (SEC) is a technique that aims to convert the expressed emotion of a spoken utterance to a target emotion while preserving the lexical and the speaker information [6,8].…”

“…While speech generation technologies have been making significant progress [4,5], emotionally expressive speech generation is still a challenge [6,7]. Speech emotion conversion (SEC) is a technique that aims to convert the expressed emotion of a spoken utterance to a target emotion while preserving the lexical and the speaker information [6,8]. Therefore, SEC has a crucial application in building next-generation human-machine interaction systems, aiming at equipping them with the ability to interact with social and emotional intelligence.…”

“…SEC datasets are primarily acted-out [7,10,15,16], as opposed to in-the-wild improvised datasets [12]. Actedout databases make a strong assumption on the availability of parallel utterances [6], i.e., one where each source utterance also has a ground-truth utterance of a target emotion, which is resource inefficient to collect [17,18] and techniques relying on them lack scalability [6].…”

“…In this work, we specifically focus on non-parallel data. Models trained on non-parallel data are scalable to different emotion types [6], as they are not restricted by the emotion pairs trained on. However, they are also more challenging than modeling parallel data, as the problem of disentanglement arises [6,8].…”

“…Models trained on non-parallel data are scalable to different emotion types [6], as they are not restricted by the emotion pairs trained on. However, they are also more challenging than modeling parallel data, as the problem of disentanglement arises [6,8]. For SEC on non-parallel data, a disentanglement method is required to decompose the input speech signal into several constituents: emotion, lexical, and speaker information, encoded in respective latent representations.…”