From linguistic anchors to acoustic skeletons: the dual-pathway model of voice identity representation

Achieving perceptual constancy for voice identity across the speech-song continuum represents a fundamental cognitive challenge. Does the auditory system rely on precise linguistic cues to "anchor" identity, or does it utilize multi-dimensional acoustic variability to build a generalized prototype? In this study, we trained listeners to recognize voices via three modalities, i.e., reading (linguistic focus), non-lexical singing on /la/ (structural focus), and singing (dual focus). We then tested their recognition of the voices under immediate, cross-modal, and one-week delayed conditions. Experiment 1 revealed that training with lyrics (reading and singing) yielded high initial accuracy, demonstrating a rapid encoding advantage of linguistic cues. However, Experiments 2 and 3 uncovered significant asymmetries. While identities learned through singing transferred seamlessly to non-lexical singing, those learned through reading incurred substantial cross-modal processing costs and marked retrieval latencies despite remaining above chance level. After a one-week delay, testing the reading group on non-lexical vocable singing revealed a behavioral dissociation. While identification accuracy remained well above chance (M = 0.62), response times spiked dramatically to 4722 ms. This latency reflects an active, exhaustive memory search triggered by persistent coarse acoustic invariants when fine-grained lexical and phonetic anchors are missing. These findings refine a Dual-Pathway Model of voice processing in which a linguistic-phonetic pathway provides language-bound precision, while an acoustic-melodic pathway exploits acoustic variability to extract a domain-general acoustic skeleton that is robust against cross-modal transfer and temporal decay.