You are viewing the SNL 2017 Archive Website. For the latest information, see the Current Website.

 
Poster C1, Thursday, November 9, 10:00 – 11:15 am, Harborview and Loch Raven Ballrooms

The Effects of Background Noise on Native and Non-native Spoken Word Recognition: An Artificial Neural Network Modelling Approach

Themis Karaminis1, Florian Hintz2, Odette Scharenborg1,3;1Centre for Language Studies, Radboud University, Nijmegen, the Netherlands;, 2Max Planck Institute for Psycholinguistics, Nijmegen, the Netherlands, 3Donders Institute for Brain, Cognition, & Behavior, Radboud University Nijmegen the Netherlands

Introduction. How does the presence of background noise affect spoken word recognition? And (how) do these effects differ when one listens to a native compared to a non-native language? We address these questions using the artificial neural-network modelling framework. We implement a series of neural network architectures, which acquire ‘native’ and ‘non-native’ lexicons of English and Dutch words. We use these architectures to simulate two spoken word recognition experiments and to systematically investigate the effects of background noise on native and non-native spoken word recognition. Method. Our model is based on ‘deep’ autoencoder neural network, which is trained on composite representations of words, consisting of phonological forms (feature-based) and their meanings. The training set comprises 121 English and Dutch translation equivalents compiled from stimuli in Scharenborg et al. (in press). Exposure to the English and the Dutch lexicons is varied so as to implement ‘native English/monolingual’ (trained only on English), ‘native English/non-native Dutch’ (English : Dutch = 3 : 1), ‘native Dutch/non-native English’ (English : Dutch = 1 : 3), and ‘native Dutch/monolingual’ (Dutch only) versions of the model. Using ‘native English’ and ‘non-native English’ versions, we simulate human performance (accuracy rates, number of erroneous responses per incorrectly identified word) in an offline spoken-word recognition experiment (Scharenborg et al., in press), in which English and Dutch students listened to English words masked with background noise, either word-initially or word-finally. Using the ‘native Dutch’ version of the model, we simulate looking preferences in an online visual-world paradigm (Hintz & Scharenborg, 2016). In this experiment, Dutch participants listened to Dutch target words while attending to pictures of the target words, phonological onset competitors of these, or words semantically related to the onset competitors (each presented along with unrelated distractors). Results. The model captures several characteristics of human performance in the two experiments. In offline spoken word recognition, the presence of background noise causes accuracy rates and the number of different erroneous responses to decrease/increase, respectively. This effect is more pronounced in word-initial compared to word-final masking conditions and in higher levels of noise, and holds for both native- and non-native listening, despite the fact that accuracy/the number of different erroneous responses are generally lower/ higher in non-native listening. In the online visual world paradigm, the presence of background noise attenuates and delays looking biases to onset phonological competitors of target words and words semantically related to the onset phonological competitors in native listening. Simulations with ‘non-native Dutch’ versions of the model predict that noise affects similarly the looking biases of non-native listeners (which are also less strong than the looking biases of native listeners). Novel data from non-native Dutch listeners completing the visual word experiment are consistent with this prediction. Conclusion. Our model unifies a wide range of empirical effects in native and non-native spoken word recognition. Our simulations suggest that the effects of noise on word recognition are largely similar in native and non-native listening and can be accounted for within the same cognitive architecture for spoken word recognition.

Topic Area: Computational Approaches

Back to Poster Schedule