Abstract
Speech emotion recognition (SER) often faces challenges due to the lack of large, annotated datasets. The presence of abundance unlabeled data offers a chance to explore methods that could significantly improve SER systems. This study explores the feasibility of enhancing general speech models by incorporating unimodal and multimodal training objectives derived from unlabeled data, specifically tailored to extract emotional content. These multimodal objectives aim to refine self-supervised learning (SSL)-based representations that, while effective in SER, were not originally created to extract emotional cues from speech. Our methodology introduces a set of multimodal objectives focused on capturing information from three primary sources: acoustic signals, through a representation objective based on the extended Geneva Minimalistic Acoustic Parameter Set (eGEMAPS); facial expressions, via visual representations obtained from a pre-trained facial expression recognition system; and textual content, through pseudo-labels generated by a pre-trained emotion sentiment model. These objectives are automatically generated from 70.7 hours of unlabeled emotional content captured in naturalistic settings. We apply our strategy to four state-of-the-art SSL-based speech models, aiming to enhance their capabilities in SER tasks with multimodal signals while still keeping inference strictly audio-only. Our experimental evaluations across the CREMA-D, MSP-IMPROV, and MSP-Podcast datasets demonstrate that our approach significantly improves SER performance, especially in settings with limited labeled data.