Answers are grounded in this paper and cite their sources.
In the past few years, self-supervised learning has revolutionalized automatic speech recognition. Self-supervised models such as wav2vec2, due to their generalization ability on huge unannotated audio corpora, were claimed to be state-of-the-art feature extractors in paralinguistic and pathological applications as well. In this study we test embeddings extracted from a wav2vec 2.0 model fine-tuned on the target language as features on a multiple sclerosis audio corpus, using three speech tasks. After comparing the resulting classification performances with traditional features such as ComParE functionals, ECAPA-TDNN and activations of a HMM/DNN hybrid acoustic model, we found that wav2vec2-based models, surprisingly, only produced a mediocre classification performance. In contrast, the decade-old ComParE functionals feature set consistently led to high scores. Our results also indicate that the number of features correlates surprisingly well with classification performance.
No takes yet. Share an insight, caveat, or question.
Gosztolya et al. (2024) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: