Multi-level Fusion of Fisher Vector Encoded BERT and Wav2vec 2.0 Embeddings for Native Language Identification

Abstract

Native Language Identification is a prominent paralinguistic study with applications ranging from biometric analysis to speaker adaptation. Former studies on this task have benefited from alternative acoustic feature representations and pre-trained neural networks. In this work, we explore the Native Language Identification performance of contextual acoustic (wav2vec 2.0) and linguistic (BERT) embeddings as state-of-the-art feature representations and combine them with acoustic features at different levels. We encode acoustic and linguistic features using Fisher Vectors, applying Fisher Vector encoding on BERT word embeddings and wav2vec 2.0 for the first time for a paralinguistic task. We compare this approach with conventional functional summarization. In line with our former study using only acoustic modality, the results indicate the superiority of Fisher Vectors encoding over the traditional techniques. Moreover, we show the efficacy of combining alternative representations now in both acoustic and linguistic modalities. Results indicate a notable contribution of the transformer-based contextual auditory and linguistic feature representations to bimodal Native Language Identification systems.