Effects of speech register and cross-linguistic variation on speech segmentation with statistical cues
Statistical learning in language acquisition was originally conceived as a gateway to speech segmentation in the absence of pre-existing knowledge about the language to be acquired. Initial support for this hypothesis came from corpus analyses that showed that Transitional Probabilities (TPs) tend to be higher within words than at word boundaries and that adults and infants can segment continuous streams of stimuli in various perceptual domains by calculating TPs (Saffran et al., 1999; Kirkham et al., 2002; Pelucchi et al., 2009). However, recent corpus studies have questioned the universality of TPs by showing that the most successful statistical segmentation strategy for a given language depends on the specific language under question. While Italian child-directed speech (Verb-Object language) is best segmented with Forward-TPs (FTPs), Hungarian (OV language) is best segmented with Backward-TPs (BTPs) (Gervain & Guevara, 2012; Saksida et al., 2016). This suggests that speech segmentation using TPs may depend on language-specific knowledge. Here we argue that TPs in the linguistic input do not only depend on the specific language to be acquired, but also on the age of the child, the input is directed to. We argue that
linguistic input to infants (IDS) or children (CDS) differs from that of adults (ADS) at the prosodic, phonological and syntactic level (Cristia, 2013).
To investigate how speech register influences statistical regularities in the input, we compared IDS and CDS from 7 languages (German, Estonian, Italian, Dutch, English, Greek, Hungarian) to children of varying ages (range 1.5 to 75 months). We compared the segmentation performance of FTPs and BTPs using a relative algorithm (word boundaries are assigned when TPs between two syllables are lower than TPs of surrounding syllable pairs) and an absolute algorithm (word boundaries are assigned when TPs drop below a fixed language-specific threshold). We find that while the type/token ratio of syllables remains constant (ß=0.019, SE=0.014, t=1.32, p=.186), the type/token ratio of words increases (ß=0.04, SE=0.012, t=3.04, p <.001) as children become older. The increase of different word types as children become older increases the number of ways existing syllable types are combined and decreases the overall value of TPs (ß=-0.4, SE=0.1, t=-7.23, p< .001). Furthermore, the performance (i.e., F-score of segmented words) of the absolute algorithm increases in language-specific ways. In Object-Verb languages (Dutch, Hungarian) segmentation with BTPs improves with age, outperforming FTPs as children get older and in Verb-Object languages (English, Estonian, Italian, Greek) FTPs improve with age outperforming BTPs as children get older. The results show that the statistical make-up of linguistic input changes as children get older, with better segmentation observed with language-specific and not universal statistical regularities.