EmoRES-TTS: Residual-Enhanced Vector Steering for Emotional Speech Generation
EmoRES steers emotion in frozen TTS models by separating a shared “non-neutral” signal from the emotion-specific residual.
The paper says this avoids retraining while improving control over requested emotions. Tested on IEMOCAP with IndexTTS-2 and CosyVoice2, EmoRES beat CoCoEmo across four objective emotion metrics. Human listeners also identified the intended dominant emotion more often, with reported gains in fidelity and naturalness preference. Work was done at Meta, and the authors say code is available. HF Daily Papers' note
The paper says this avoids retraining while improving control over requested emotions. Tested on IEMOCAP with IndexTTS-2 and CosyVoice2, EmoRES beat CoCoEmo across four objective emotion metrics. Human listeners also identified the intended dominant emotion more often, with reported gains in fidelity and naturalness preference. Work was done at Meta, and the authors say code is available. HF Daily Papers' note
score 4