Evidence of Vocal Tract Articulation in Self-Supervised Learning of Speech

Numerous self-supervised learning (SSL) models for speech have been proposed for pre-training models of speech representations, and recent SSL models are very successful in diverse downstream tasks. To understand such utilities, previous works probe representations of speech models to reveal which & how speech related information is encoded in the learned representations. While encoding properties have been extensively explored from the perspective of acoustics, phonetics, and semantics, the physical grounding by speech production has not yet received full attention. To bridge this gap, we conduct a comprehensive analysis to link speech representations to articulatory trajectories measured by electromagnetic articulography (EMA). Our analysis is based on a linear probing approach where we measure articulatory score as an average correlation of linear mapping to EMA. We analyze a set of SSL models selected from the leaderboard of the SU- PERB benchmark and perform further detailed analyses on two major models, Wav2Vec 2.0 and HuBERT. Surprisingly, representations from the recent speech SSL models are highly correlated with EMA traces (best: r = 0.81), and only 5 minutes were sufficient to train a linear model with high performance (r = 0.77). Our findings suggest that SSL models learn to closely align with continuous articulations and provide a novel insight into speech SSL.