Probing speech emotion recognition transformers for linguistic knowledge

A Triantafyllopoulos, J Wagner, H Wierstorf, M Schmitt, U Reichel, F Eyben, F Burkhardt, BW Schuller, "Probing speech emotion recognition transformers for linguistic knowledge," in Proceedings Interspeech 2022, (2022). [ paper ]

Bibtex

@proceedings{Triantafyllopoulos2022a,
    title     = {Probing speech emotion recognition transformers
                 for linguistic knowledge},
    author    = {Triantafyllopoulos, Andreas and Wagner, Johannes
                 and Wierstorf, Hagen and Schmitt, Maximilian
                 and Reichel, Uwe, and Eyben, Florian
                 and Burkhardt, Felix, and Schuller, Björn W.},
    booktitle = {Proceedings Interspeech 2022},
    publisher = {ICSA},
    address   = {Incheon, Korea},
    month     = {September},
    pages     = {146--150},
    doi       = {10.21437/Interspeech.2022-10371},
    year      = {2022}
}

Abstract

Large, pre-trained neural networks consisting of self-attention layers (transformers) have recently achieved state-of-the-art results on several speech emotion recognition (SER) datasets. These models are typically pre-trained in self-supervised manner with the goal to improve automatic speech recognition performance – and thus, to understand linguistic information. In this work, we investigate the extent in which this information is exploited during SER fine-tuning. Using a reproducible methodology based on open-source tools, we synthesise prosodically neutral speech utterances while varying the sentiment of the text. Valence predictions of the transformer model are very reactive to positive and negative sentiment content, as well as negations, but not to intensifiers or reducers, while none of those linguistic features impact arousal or dominance. These findings show that transformers can successfully leverage linguistic information to improve their valence predictions, and that linguistic analysis should be included in their testing.