I-vectors and Deep Convolutional Neural Networks for Language Identification in Clean and Reverberant Environments

Abstract

AbstractIn the current study, a method for automatic language identification based on deep convolutional neural networks (DCNN) and the i-vector paradigm is proposed. Convolutional neural networks (CNN) have been successfully applied to image classification, speech emotion recognition, and facial expression recognition. In the current study, a variant of typical CNN is being applied and experimentally investigated in spoken language identification. When the proposed method was evaluated on the NIST 2015 i-vector Machine Learning Challenge task for the recognition of 50 in-set languages, a 3.9% equal error rate (EER) was achieved. The proposed method was compared to two baseline methods showing superior performance. The results obtained are very promising and show the effectiveness of using DCNN in spoken language identification. Furthermore, in the current study, a front-end feature enhancement and dereverberation approach based on a deep convolutional autoencoder is also reported.KeywordsSpoken language identificationDeep convolutional neural networksDereverberationDenoising autoencoder