Detecting the native language (L1) of non-native English speakers may be of great relevance in some applications, such as computer assisted language learning or IVR services. In fact, the L1 detection problem closely resembles the problem of spoken language and dialect recognition. In particular, log-likelihood ratios of phone posterior probabilities, known as Phone LogLikelihood Ratios (PLLR), have been recently introduced as features for spoken language recognition systems. This representation has proven to be an effective way of retrieving acoustic-phonotactic information at frame-level, which allows for its use in state-of-the-art systems, that is, in i-vector systems. In this paper, we explore the use of PLLR-based i-vector systems for L1 native language detection. We also investigate several linear and non-linear L1 classification schemes on top of the PLLR i-vector front-ends. Moreover, we compare PLLR based systems with both conventional phonotactic systems based on n-gram modelling of phoneme sequences and acoustic-based i-vector systems. Finally, the potential complementarity of the different approaches is investigated based on a set of system fusion experiments.
Cite as: Abad, A., Ribeiro, E., Kepler, F., Astudillo, R., Trancoso, I. (2016) Exploiting Phone Log-Likelihood Ratio Features for the Detection of the Native Language of Non-Native English Speakers. Proc. Interspeech 2016, 2413-2417, doi: 10.21437/Interspeech.2016-1491
@inproceedings{abad16_interspeech,
author={Alberto Abad and Eugénio Ribeiro and Fábio Kepler and Ramon Astudillo and Isabel Trancoso},
title={{Exploiting Phone Log-Likelihood Ratio Features for the Detection of the Native Language of Non-Native English Speakers}},
year=2016,
booktitle={Proc. Interspeech 2016},
pages={2413--2417},
doi={10.21437/Interspeech.2016-1491}
}