MTLSER: Multi-task learning enhanced speech emotion recognition with pre-trained acoustic model

This study proposes a novel Speech Emotion Recognition (SER) approach employing a Multi-Task Learning framework (MTLSER), designed to boost recognition accuracy by training multiple related tasks simultaneously and sharing information via a joint loss function. This framework integrates SER as the primary task, with Automatic Speech Recognition (ASR) and speaker identification serving as auxiliary tasks. Feature extraction is conducted using the pre-trained wav2vec2.0 model, which acts as a shared layer within our multi-task learning (MTL) framework. Extracted features are then processed in parallel by the three tasks. The contributions of auxiliary tasks are adjusted through hyperparameters, and their loss functions are amalgamated into a singular joint loss function for effective backpropagation. This optimization refines the model's internal parameters. Our method's efficacy is tested during the inference stage, where the model concurrently outputs the emotion, textual content, and speaker identity from the input audio. We conducted ablation studies and a sensitivity analysis on the hyperparameters to determine the optimal settings for emotion recognition. The performance of our proposed MTLSER model is evaluated using the public IEMOCAP dataset. Results from extensive testing show a significant improvement over traditional methods, achieving a Weighted Accuracy (WA) of 82.63% and an Unweighted Accuracy (UA) of 82.19%. These findings affirm the effectiveness and robustness of our approach. Our code is publicly available at https://github.com/CCNU-nercel-lc/MTL-SER.