JKIIE 2021 · First author · Dec 2021
Multi-modal Korean Emotion Recognition with Consistency Regularization
01Background & Goals
- We studied a multimodal framework that adds an external speech-recognition module to improve speech-based emotion recognition.
- Real-world speech recognition makes mistakes, so we needed a training method robust to them.
02Contributions
Multimodal framework and robust training
- Designed a multimodal architecture over speech and STT text (BERT, Wav2vec 2.0)
- Designed a consistency-regularization loss that aligns the prediction distributions of the gold-text and STT paths
03Tech Stack
- Framework / Platform
- PyTorch, Fairseq, Hydra
- Methodology
- Multimodal learning, Self-supervised learning, Consistency regularization
04Results
- Published in JKIIE 47(6) (first author); code released
- Best Master's Thesis Award, KIIE Master's Thesis Competition (Nov 2021)