Reproduce “Speaker-agnostic Emotion Vector for Cross-speaker Emotion Intensity Control” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Masato Murata, Koichi Miyazaki, Tomoki Koriyama
This paper from CyberAgent introduces a speaker-agnostic emotion vector for cross-speaker emotion intensity control in speech synthesis, extending the "emotion arithmetic" paradigm to generalize across diverse speakers. The method enables high-quality emotional speech synthesis for unseen speakers, preserving speaker identity and providing perceivable control over emotion intensity, with MOS scores for naturalness comparable to same-speaker generation in cross-speaker settings.
Key insights
- Training a multi-speaker TTS model and deriving an emotion vector from a multi-speaker context captures shared, speaker-independent emotional expressions, overcoming limitations of single-speaker approaches.
- Parameter-space manipulation (task arithmetic) can be effectively extended for cross-speaker generalization of complex speech attributes like emotion by incorporating robust speaker conditioning.
- Speaker embeddings (x-vectors) are crucial for maintaining speaker consistency when transferring emotions to different speakers, including those not seen during training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.