반응형
[Paper 리뷰] Emotion2Vec: Self-Supervised Pre-Training for Speech Emotion Representation
Emotion2Vec: Self-Supervised Pre-Training for Speech Emotion RepresentationUniversal speech emotion representation이 필요함Emotion2VecSelf-Supervised Online Distillation을 통해 unlabled emotion data로 pre-trainingPre-training 시 utterance-level loss와 frame-level loss를 combine논문 (ACL 2024) : Paper Link1. IntroductionSpeech에서 emotion을 추출하기 위해서는 주로 Filter Bank (FBank)나 MFCC를 활용함BUT, 해당 feature는 rich semanti..
Paper/Representation
2025. 5. 24. 07:41
반응형