Evaluating the learnability of vowel categories from infant-directed speech

Abstract

Hyper-articulated vowels are considered a hallmark of infant-directed Speech (IDS) and are thought to facilitate acquisition. Consistent with this idea, vowels in IDS are reported to be more peripheral than in adult-directed speech (ADS) (e.g., Trainor and Desjardins, 2002; Liu et al., 2005). However, there are also reports that vowels in IDS are more variable (Cristia and Seidl, 2014; Martin et al., 2015). We evaluated the learnability of vowel categories from IDS in two languages—American English, which has a crowded vowel space, and Mexican Spanish, which has fewer vowels. First, we used k-means clustering, an unsupervised learning algorithm, to partition vowels based on spectral and duration measures from conversational IDS. The algorithm was trained on IDS from 5 English speakers (Providence Corpus, Demuth et al., 2006) and 12 Spanish speakers (Aly et al., 2016). The vowel categories extracted from IDS were used to classify ADS vowels from 5 English speakers from the Buckeye corpus and 12 Spanish speakers (Kim and Repiso Puigderlliura, 2021). Next, we switched the training and testing set, by training on ADS, and testing on IDS. If IDS is facilitative, we expect a higher accuracy of vowel classification with training on IDS rather than ADS.