Learning Prosodic Focus from Continuous Speech Input: A Neural Network Exploration

This study uses connectionist modeling to explore whether and how infants might learn prosodic focus directly from continuous speech input. Focus is a communicative function that serves to put emphasis on a particular part of an utterance, and it is mainly encoded by pitch variations. The acquisition of focus entails two major difficulties. The first is that focus-related pitch patterns are confounded by other linguistic functions that also use pitch for their encoding, such as lexical tone in a tone language. Second, speakers have different pitch ranges, which further confounds the focus related pitch patterns. In three simulations using self-organizing neural networks, we explored how focus may be learned from continuous acoustic signals in Mandarin that were produced with cooccurring lexical tones and by multiple speakers. We used sentence-sized F0 contours as well as their velocity profiles (D1) as training input. Results show that both F0 and D1 contours provide information for focus learning, but only the D1-trained network adequately handled the variability introduced by cross-gender differences. The recognition rate was analogous to human performance. Implications of these findings for theories of language acquisition and adult speech perception are discussed.