L1 Identification from L2 Speech Using Neural Spectrogram Analysis

Abstract

The main objective of this project is to model L1-L2 interaction and uncover discriminative speech features that can identify the L1 background of a speaker from their non-native English speech. Traditional phonetic analyses of L1-L2 interaction tend to use a pre-selected set of acoustic features. This, however, may not be sufficient to capture all traces of the L1 in the L2 speech to make an accurate classification. Deep learning has the potential to address this by exploring the space of features automatically. In this talk I report a series of classification experiments involving a deep convolutional neural network (CNN) based on spectrogram pictures. The classification problem consists of determining whether English speech samples from a large spontaneous speech corpus are spoken by a native speaker of SSBE, Japanese, Dutch, French or Polish.