CNN-Based End-To-End Language Identification

Abstract

Recently, language identification (LID) on long utterances has archived very low error rate, however, it is still a challenging task under short-duration condition. In this paper, we propose an end-to-end short-duration language identification system based on deep convolutional neural network (DCNN), where the whole network is trained with multi-class cross-entroy loss. Besides, we compare three kinds of input features: Mel-Frequency Cepstral Coefficients (MFCC), log Mel-scale Filter Bank energies (FBANK) and spectrogram energies. The experimental results indicate that spectrogram energies achieves the best performance among them In order to enhance the robustness of system, at the training stage, the databases are augmented by applying time-scale modification (TSM) method. Based on APl 7-OLR databases, under 1-second condition, the proposed system has improved 32.7% than traditional i-vector system, and compared with other neural network systems, it peforms equally well and even better.