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

language recognition

In this work, we analyze different designs of a language identification (LID) system based on embeddings. In our case, an embedding represents a whole utterance (or a speech segment of variable duration) as a fixed-length vector (similar to the ivector). Moreover, this embedding aims to capture information relevant to the target task (LID), and it is obtained by training a deep 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, whose frame-by-frame outputs are summarized into mean and standard deviation statistics for each utterance. After this pooling layer, we add two fully connected layers whose outputs are used as embeddings, which are afterwards modeled by a Gaussian linear classifier (GLC). For training, we add a softmax output layer and train the whole network with multi-class cross-entropy objective to discriminate between languages. We analyze the effect of using data augmentation in the DNN training, as well as different input features and architecture hyper-parameters, obtaining configurations that gradually improved the performance of the embedding system. We report our results on the NIST LRE 2017 evaluation dataset and compare the performance of embeddings with a reference i-vector system. We show that the best configuration of our embedding system outperforms the strong reference i-vector system by 3% relative, and this is further pushed up to 10% relative improvement via a simple score level fusion.