Word stress in self-supervised speech models: A cross-linguistic comparison
This work addresses the problem of understanding linguistic feature learning in speech models for researchers in computational linguistics and speech technology, but it is incremental as it applies existing methods to new data.
The study investigated word stress representations in self-supervised speech models (Wav2vec 2.0) across five languages, finding that diagnostic classifiers could distinguish stressed and unstressed syllables with high accuracy, and that representations were language-specific, with greater differences between variable-stress and fixed-stress languages.
In this paper we study word stress representations learned by self-supervised speech models (S3M), specifically the Wav2vec 2.0 model. We investigate the S3M representations of word stress for five different languages: Three languages with variable or lexical stress (Dutch, English and German) and two languages with fixed or demarcative stress (Hungarian and Polish). We train diagnostic stress classifiers on S3M embeddings and show that they can distinguish between stressed and unstressed syllables in read-aloud short sentences with high accuracy. We also tested language-specificity effects of S3M word stress. The results indicate that the word stress representations are language-specific, with a greater difference between the set of variable versus the set of fixed stressed languages.