Learning phonemic vowel length from naturalistic recordings of Japanese infant-directed speech

Abstract

In Japanese, vowel duration can distinguish the meaning of words. In order for infants to learn this phonemic contrast using simple distributional analyses, there should be reliable differences in the duration of short and long vowels, and the frequency distribution of vowels must make these differences salient enough in the input. In this study, we evaluate these requirements of phonemic learning by analyzing the duration of vowels from over 11 hours of Japanese infant-directed speech. We found that long vowels are substantially longer than short vowels in the input directed to infants, for each of the five oral vowels. However, we also found that learning phonemic length from the overall distribution of vowel duration is not going to be easy for a simple distributional learner, because of the large base-rate effect (i.e., 94% of vowels are short), and because of the many factors that influence vowel duration (e.g., intonational phrase boundaries, word boundaries, and vowel height). Therefore, a successful learner would need to take into account additional factors such as prosodic and lexical cues in order to discover that duration can contrast the meaning of words in Japanese. These findings highlight the importance of taking into account the naturalistic distributions of lexicons and acoustic cues when modeling early phonemic learning.

Figure 1. Mean duration of short and long vowels in the present Japanese IDS corpus.

The difference in duration between short and long vowels is reliable and the effect size is large. The error bars represent the standard error of the mean for each vowel across participants.

Figure 2. The frequency distribution of vowel duration in the present Japanese IDS corpus.

Ninety-four percent of the vowels in our corpus are short, and there is complete overlap in the distribution of short and long vowels. This kind of input is problematic for simple distributional learning models.