Transfer Learning of Large Speech Models for Italian Speech Emotion Recognition

Recent research in Automated Speech Recognition (ASR) has shifted towards using large pre-trained speech models trained on extensive corpora with a Self-Supervised Learning (SSL) approach. These models can transfer general-purpose knowledge to tasks like Speech Emotion Recognition (SER). Due to their highly parameterized architecture, fine-tuning all the weights is computationally inefficient. Consequently, new Parameter-Efficient Fine-Tuning (PEFT) strategies have been explored for the SER task in English. Given the lack of SSL speech models in Italian, current models are either English-only or multilingual, with little effort made to adapt them to SER in Italian. In this work, we investigate transfer learning performance on Italian SER using PEFT strategies, marking the first exploration in this direction. We apply PEFT techniques, such as Low-Rank Adaptation (LoRA) and Adapter, on Italian SER datasets Emozionalmente, DEMoS, and EMOVO. Results show LoRA is the most effective PEFT technique for Italian SER. Speech models pre-trained on large-scale English corpora perform comparably to, or better than, multilingual ones, even when specialized in Italian before the SER task, suggesting some shared paralinguistic features between the languages.