Speech-Based Screening of Multiple Sclerosis By Features Derived from Self-Supervised Models

Multiple sclerosis (MS) is a chronic inflammatory disease of the central nervous system. Since it, among other symptoms, adversely affects the speech of the subject, automatic speech analysis might offer a simple, inexpensive and remote tool for MS screening or monitoring the progression of the disease. We employ ten different wav2vec 2.0 models as the base of feature extraction and compare the performance with pre-trained and custom x-vector models. Based on our results, cross-lingual models perform better than the base wav2vec 2.0 networks, but the model size is crucial as the best results were obtained with a model having one billion trainable parameters. We found fine-tuning the application language to be beneficial to the classification performance, but for other languages, it did not improve the AUC scores. Surprisingly, though, we did not outperform standard x-vectors, which might be due to the standard, but perhaps too simple aggregation strategy of the frame-level embeddings.