Spoken Language Identification with Deep Convolutional Neural Network and Data Augmentation

In this paper, a spoken language detection system based on deep convolutional neural networks is presented. The neural network model is trained and tested on a speech dataset containing five languages. Speech signals are first converted into mel-spectrogram features and these features are fed into the deep convolutional neural network. Flattened outputs of the deep convolutional network are then fed into a recurrent layer, and a dense layer with softmax activation function is used as an output layer to predict the output language probabilities. This network results in 0.89 F1-score in our test data. We also used a data augmentation method, namely Spec Augment, which increased the F1-score to 0.94.