Multi-Lingual Multi-Task Speech Emotion Recognition Using wav2vec 2.0

Speech Emotion Recognition (SER) has several use cases for Digital Entertainment Content (DEC) in Over-the-top (OTT) services, emotive Text-to-Speech (TTS) engines and voice assistants. In this work, we present a Multi-Lingual (MLi) and Multi-Task Learning (MTL) audio only SER system based on the multi-lingual pre-trained wav2vec 2.0 model. The model is fine-tuned on 25 open source datasets in 13 locales across 7 emotion categories. We show that, a) Our wav2vec 2.0 single task based model outperforms Pre-trained Audio Neural Network (PANN) based single task pre-trained model by 7.2% (relative), b) The best MTL model outperforms the PANN based and wav2vec 2.0 based single task models by 8.6% and 1.7% (relative) respectively, c) The MTL based system outperforms pre-trained single task wav2vec 2.0 model in 9 out of 13 locales in terms of weighted F1 scores, and d) The MTL-MLi wav2vec 2.0 outperforms the state-of-the-art for the languages contained in the pre-training corpora.