Artificial intelligence has made significant progress in understanding human emotions. Earlier this month, the second Multimodal Emotion Recognition Challenge (MER24) successfully concluded. This high-profile event, initiated by several international renowned scholars, aims to promote the application of AI emotion recognition technology in real human-machine interaction scenarios.
MER24 features three tracks, with the Semi track garnering significant attention due to its high difficulty and intense competition. The Semi track requires participating teams to train models using a small amount of labeled and a large amount of unlabeled video data, and to evaluate the model's performance and generalization capabilities on unlabeled datasets. The voice technology team from Soul App won first place in this track by leveraging innovative technical solutions.
Competition website: https://zeroqiaoba.github.io/MER2024-website/#organization
The success of the Soul team is attributed to its deep accumulation and innovation in multimodal data understanding, emotion recognition algorithms, model optimization platform tools, internal workflow construction, and efficient teamwork. Facing the challenge of scarce data, the Soul team adopted various strategies, including improving semi-supervised learning techniques, leveraging pretrained models to extract multimodal features, proposing effective feature fusion methods, and innovative models for video and text modalities.
The technical solution of the Soul team not only enhanced the accuracy of emotion recognition but also better differentiated the boundaries of easily confused emotions. This achievement is a concentrated reflection of Soul's deep cultivation of AI large model technology, especially its multimodal emotional interaction capabilities in the social field.
The demand for emotional AI in the social field is growing. Soul has transformed from a "task executor" to a "companion fulfilling human emotional needs" by building an AI with emotional capabilities. Products developed by Soul, such as AI Gou Dan, Werewolf Enchantment Game, and Echoes of Another World application, demonstrate Soul's integration capabilities in anthropomorphism, knowledge, multimodality, and time perception, providing users with rich and warm AI interactive experiences.
2024 is considered the元年 of AIGC applications, and domestic companies like Soul have achieved significant results in the AI social direction through self-developed technology accumulation. Soul has incubated a series of products based on its self-developed language and voice large models, and has accumulated rich innovative technologies and practical experiences in enhancing AI's emotional interaction with users.
In the future, platforms like Soul that persist in technological and product innovation will continue to create value for users, achieving more enduring and diverse commercial value on the basis of forming a thriving content and community ecosystem.