Project 6: Low-level processing in L2 speech perception
Speech recognition by L2 learners suffers in adverse listening conditions. Several factors are likely to be involved in this performance disparity: (i) interference from the native language; (ii) incomplete acquisition (lack of FPD) of L2 categories; and (iii) presumed universal confusions due to inherent maskability of certain sounds. It is important to tease apart these factors if effective L2 training regimes are to be devised. In addition, knowledge of L2 perception will lead to insights into processes and representations used by native listeners. The purpose of this study is to clarify the role of the above factors via a tightly-linked series of behavioural and computer modelling studies.
Perception To assess the role of L1 interference in L2 perception, common speech perception tasks will be carried out in a range of language communities. In one set of experiments, VCV tokens will be used. Data will be collected from native language groups as well as L2 learners. Amongst other contrasts, we will study the different status and FPD of the interdental fricatives (i.e. the initial sound in the English words “this” and “think”) in Spanish, Czech and English, and their acquisition by learners in each language group. Results will feed into a computational modelling study described below. A parallel thread will investigate L2 learning of subsegmental lexical information affected by assimilatory processes (e.g. when bad girl sounds like bag girl; and voicing assimilation in French). Studies of disambiguation of such assimilated sequences by native speakers suggest that FPD must be part of the lexical representation of the word, mediating between acoustic input and phonological representation. We will assess (1) whether L1 assimilation patterns are transferred to L2, whether appropriate or not, (2) whether L2 learners acquire assimilatory contingencies as L1 learners do, i.e. implicitly.
Computational modelling Since non-native listener populations vary widely in the degree and type of exposure to the L2, it is difficult to control for each of the factors which influence L2 speech perception. One solution is to construct a computer model in which L2 ‘exposure’ can be varied continuously. A model can also shed light on the relative roles of native language interference and incomplete acquisition. A computational model will be constructed using standard ASR techniques. To model a L2 learner’s capabilities, different model sets will be trained for the L1 and L2 using speech material for each of the languages used in the perception tests. By varying the amount of training material used to train the L2 model, the process of L2 acquisition will be simulated. For speech in noise data, a missing data ASR model developed at Sheffield will be used to model energetic masking. Consequently, the effect of adverse conditions on L2 speech perception will be tested to analyse the robustness of acquired representations. This modelling study will not only lead to a better understanding of early processes in L2 perception, but may cast light on the reasons for robustness of native perception.
Young researchers One ESR (Sankowska) is based in the Basque Country. There will also be an ER based in Prague from mid-2009. Between these two Fellows, they will make visits to up to 5 other countries for perception testing with L2 learners. They can visits Sheffield, Leuven and/or Trondheim for computer modelling. This project links linguistics, phonetics, psychology and computer science.
Links L2 issues permeate all themes. In particular, Projects 1, 3 and 4 involve cross-linguistic comparisons and data, with Project 4 also addressing assimilation.
Working on this project: » Dr M. Luisa García Lecumberri » Prof Dr Mircea Giurgiu » Dr Jan Volín » Joanna Sankowska » Prof Martin Cooke » Prof Jasone Cenoz » Dr Radek Skarnitzl » Dr Maria Paola Bissiri
