Evidence of predictive information compression in latent space in humans during speech listening

Speech perception requires transforming acoustic input into neural representations that support linguistic understanding, yet its underlying computational principles remain unclear. Classical efficient coding theories posit optimal compression of sensory input, whereas alternative accounts propose that neural systems preferentially encode information that supports prediction. A key open question is whether such predictive encoding operates on fixed inputs or on flexible internal representations. We instantiated three hypothesis models of speech processing: (i) optimal compression with deep autoencoders, (ii) predictive reconstruction with predictive autoencoders, and (iii) predictive information representation via latent-space prediction using contrastive learning. We compared resulting speech latent representations to electroencephalographic (EEG) activity during speech listening. Representations learned under the predictive information objective best explained neural latents. Crucially, only representations that selectively compressed predictive information predicted behavioral performance, suggesting that neural speech representations are structured to encode predictive information in latent space rather than to maximize compression or input prediction.