Surface forms trump underlying representations in functional generalisations in speech perception: the case of German devoiced stops

Listeners can adapt their perceptual categories for speech sounds in response to speakers' unusual pronunciations. The present study tested whether this generalisation is sensitive to surface or underlying properties of speech by exploiting the devoicing of voiced stops in German. This allows us to assess whether learning on phonetically voiceless stops that are underlyingly voiced generalises to stops that share the same surface form (i.e. voiceless) or the same underlying representation (i.e. voiced). Our results showed only minimal generalisation: learning for (surface) voiceless stops in offset position that are underlyingly voiced generalises to surface and underlying voiceless stops in the same position but neither to voiced nor voiceless stops in intervocalic position. This suggests that listeners extract segments of sufficient acoustic similarity from the input and use them for generalisation of learning in speech perception. The units of perception thereby appear context sensitive rather than abstract phonemes or phonological/articulatory features.