[b] = [d] - [t] + [p]: Self-supervised Speech Models Discover Phonological Vector Arithmetic
Self-supervised speech models (S3Ms) encode speech using vectors that behave like phonological features. The research found that these representations support linear arithmetic, meaning you can combine or scale these vectors to control aspects of speech, such as voicing or place of articulation. This shows S3Ms learn a structured, compositional way to represent sound.