DNN Based Embeddings for Language Recognition

DNN Based Embeddings for Language Recognition

Paper in proceedings (conference paper)

In this work, we present a language identification (LID) systembased on embeddings. In our case, an embedding is a fixed-lengthvector (similar to i-vector) that represents the whole utterance, butunlike i-vector it is designed to contain mostly information relevantto the target task (LID). In order to obtain these embeddings, wetrain a deep neural network (DNN) with sequence summarizationlayer to classify languages. In particular, we trained a DNN basedon bidirectional long short-term memory (BLSTM) recurrent neuralnetwork (RNN) layers, whose frame-by-frame outputs are summarizedinto mean and standard deviation statistics. After this poolinglayer, we add two fully connected layers whose outputs correspondto embeddings. Finally, we add a softmax output layer and train thewhole network with multi-class cross-entropy objective to discriminatebetween languages. We report our results on NIST LRE 2015and we compare the performance of embeddings and correspondingi-vectors both modeled by Gaussian Linear Classifier (GLC). Usingonly embeddings resulted in comparable performance to i-vectorsand by performing score-level fusion we achieved 7.3% relativeimprovement over the baseline.

In this work, we present a language identification (LID) systembased on embeddings. In our case, an embedding is a fixed-lengthvector (similar to i-vector) that represents the whole utterance, butunlike i-vector it is designed to contain mostly information relevantto the target task (LID). In order to obtain these embeddings, wetrain a deep neural network (DNN) with sequence summarizationlayer to classify languages. In particular, we trained a DNN basedon bidirectional long short-term memory (BLSTM) recurrent neuralnetwork (RNN) layers, whose frame-by-frame outputs are summarizedinto mean and standard deviation statistics. After this poolinglayer, we add two fully connected layers whose outputs correspondto embeddings. Finally, we add a softmax output layer and train thewhole network with multi-class cross-entropy objective to discriminatebetween languages. We report our results on NIST LRE 2015and we compare the performance of embeddings and correspondingi-vectors both modeled by Gaussian Linear Classifier (GLC). Usingonly embeddings resulted in comparable performance to i-vectorsand by performing score-level fusion we achieved 7.3% relativeimprovement over the baseline.

Embeddings, language recognition, LID, DNN

Embeddings, language recognition, LID, DNN

LOZANO DÍEZ, A.; PLCHOT, O.; MATĚJKA, P.; GONZALEZ-RODRIGUEZ, J.

2019

15.04.2018

IEEE Signal Processing Society

Calgary

978-1-5386-4658-8

Proceedings of ICASSP 2018

5184

5188

5

https://www.fit.vut.cz/research/publication/11723/

http://hdl.handle.net/

lozano_icassp2018_0005184