Spectrogram Enhanced SimCLRV2 Emotional Representation Strategy for Kid’s Speech using Multisource Transfer Learning in CNN
Year: 2024
Type: article
Abstract: A model that learns voice image representation, which outperform past techniques is developed. An unlabeled dataset is used to develop a semi-supervised representation contrastive learning strategy that picks and compares anchor, negative, and positive (APN) features with just 1% training data of real-time kid’s speech. Most of kid’s emotion detection model evaluate its model with adult’s emotion dataset, which may not give accurate result. To overcome the limitation of less labelled kid’s dataset we use Spectrogram Enhanced SimCLRV2 model since we can train the model with minimal available kid’s dataset with which we can classify and predict kid’s emotion effectively. Its goal is to maximize agreement across variously enhanced samples in the latent region of the input representation using contrastive loss. The quality of acquired emotional representations is greatly enhanced by the introduction of learnable nonlinear transformations between learnt emotional representations and contrastive losses. Multi-source transfer learning develops the network's capacity for accurate classification, finds the missing information in each source, and transfers it to the target data to complete it. As of our knowledge this is the first work that have utilized IMTL with SimCLR to improve target labels as well as overcomes less data. In order to train the network without labels using a self-supervised algorithm we use Ravdess song and Iemocap dataset. We use real-time kid’s data as the teacher network. We employ Ravdess speech with labels as supervised algorithm. By incorporating these discoveries and analyzing those using SimCLR CSL methodologies in Zenodo children recording un-labeled dataset, which is the student network, we can dramatically outperform previous methods for semi-supervised learning in recognition rate of kid’s emotion by training network with 1% real-time kid’s data.
Institution: SRM Institute of Science and Technology
Language: English
Topic: Emotion and Mood Recognition
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7218249172
