Human and Computer Acquisition of Vowel Categories

This paper describes work in progress on computer modeling of speech acquisition. In these tests, a computer model is presented with natural speech to examine whether the model can learn vowel categories. Preliminary results are promising. The computer learns a threevowel and a five-vowel system from consonant vowel syllables in isolation. However, the experiments show that clear input is crucial for learning categories and that the most frequent variety of speech, rapid casual speech, does not provide such a signal. We propose that motherese, the speech register that parents use for addressing their children, can provide such an input. The usefulness of motherese in learning vowel categories can be tested with the computer model using existing recordings of motherese and adult-to-adult speech. Computer modeling of this type is compared with speech recognition systems.