Reference model
Baseline
Grad-TTS with emotion embedding
Interspeech 2026 · Audio demos
Beyond
One-Size-Fits-All.
Personalized and Culturally Adaptive Emotional TTS via Interactive Optimization of Individual Emotion Perception Spaces
Nara Institute of Science and Technology, Japan
Emotion is personal.
The same emotional expression can be perceived differently across listeners and cultures. We personalize arousal–valence (A–V) coordinates through listener feedback, adapting emotional speech without retraining the TTS backbone.
01 /
Speech quality & emotional similarity
Compare Grad-TTS with an emotion embedding against the proposed emotion controller and ground-truth speech, before interactive personalization.
Headphones recommended. Playing a sample pauses the previous one.
Confused
Target emotionOur method
Proposed
Grad-TTS with emotion controller
Recorded speech
Ground truth
Reference recording
02 /
Personalized emotional expression
Listen to the averaged U.S. reference mapping alongside an individual listener’s personalized A–V mapping. Each pair demonstrates a different target emotion and listener background.
Angry
Target emotionPopulation mapping
Averaged reference
A–V values from the U.S.-based reference dataset
Individual mapping
Personalized · China
A–V values adapted to an individual listener
Sad
Target emotionPopulation mapping
Averaged reference
A–V values from the U.S.-based reference dataset
Individual mapping
Personalized · Indonesia
A–V values adapted to an individual listener
Neutral
Target emotionPopulation mapping
Averaged reference
A–V values from the U.S.-based reference dataset
Individual mapping
Personalized · Japan
A–V values adapted to an individual listener
03 /
Cross-cultural emotional expression
Explore the same target emotion with culture-specific A–V mappings. Country labels refer to participant groups; the speech samples are in English.
Happy
Target emotionParticipant group
Indonesia
Culture-specific A–V mapping
Participant group
Japan
Culture-specific A–V mapping
Participant group
China
Culture-specific A–V mapping