Analysis of DNN-based Embeddings for Language Recognition on the NIST LRE 2017

In this work, we analyze different designs of a language identification(LID) system based on embeddings. In our case, anembedding represents a whole utterance (or a speech segmentof variable duration) as a fixed-length vector (similar to the ivector).Moreover, this embedding aims to capture informationrelevant to the target task (LID), and it is obtained by training adeep neural network (DNN) to classify languages. In particular,we trained a DNN based on bidirectional long short-term memory(BLSTM) recurrent neural network (RNN) layers, whoseframe-by-frame outputs are summarized into mean and standarddeviation statistics for each utterance. After this pooling layer,we add two fully connected layers whose outputs are used asembeddings, which are afterwards modeled by a Gaussian linearclassifier (GLC). For training, we add a softmax output layerand train the whole network with multi-class cross-entropy objectiveto discriminate between languages. We analyze the effectof using data augmentation in the DNN training, as well asdifferent input features and architecture hyper-parameters, obtainingconfigurations that gradually improved the performanceof the embedding system. We report our results on the NISTLRE 2017 evaluation dataset and compare the performance ofembeddings with a reference i-vector system. We show thatthe best configuration of our embedding system outperforms thestrong reference i-vector system by 3% relative, and this is furtherpushed up to 10% relative improvement via a simple scorelevel fusion.