Infants learn to imitate and recognize words at an early age,but phonemic awareness develops at a later age, guided byacquisition of literacy for example. We investigate ahypothesis that speech representations in the brain are formedpartly due to articulatory-acoustic learning, and theserepresentations may be used as a basis when learning anadditional mapping to phonemes. We train a convolutionalrecurrent neural network, having an articulatory branch and aphonemic branch for multitask learning. When trained withreal conversational speech and aligned synthesizedarticulation, it is shown that the use of the articulatoryrepresentation boosts phoneme recognition accuracy, whenthe first convolutional layers are shared between the twobranches. It is hypothesized that representations involved inspeech perception formed in the brain during childhood maybe partly based on articulatory learning, and an additionalmapping from these low-level speech representations tophonemes has to be learned.

Existing models of infant word learning have mainly assumed that the learner is capable of segmenting words from speech before grounding them to their referential meaning, while segmentation itself has been treated relatively independently of meaning acquisition. In this paper, we argue that situated cues such as visually perceived concrete objects or actions are not just important for word-to-meaning mapping, but that they are useful in pre-linguistic word segmentation, thereby helping the learner to bootstrap the language learning process. We present a model where joint acquisition of proto-lexical segments and their meanings maximizes the referential quality of the lexicon, and where learning can occur without any a priori knowledge of the language or its linguistically relevant units. We investigate the behavior of the model using a computational implementation of statistical learning, showing successful word segmentation under varying degrees of referential uncertainty.

Human infants learn meanings for words in interaction with their environment. Individual learning scenarios can be ambiguous due to the presence of several words and possible meanings. One possible way to overcome ambiguity is called cross-situational learning (XSL), where information is gathered over several learning trials. Experimental studies of human XSL have shown that cognitive constraints, such as attention and memory limitations, decrease human performance when compared to computer models that can store all available information. In this paper, we approach modeling of human performance with a novel computational XSL algorithm, FAMM (Familiarity preference, Associative learning, Mutual exclusivity, Memory decay), equipped with the four main components motivated by experimental research. The model is evaluated based on a number of earlier XSL experiments that probe different aspects of learning. FAMM is shown to provide a better fit to the behavioral data than the earlier proposed model of Kachergis et al. (2012