关键词: 语种识别, 融合特征, 深度神经网络, 低信噪比, 鲁棒性

Abstract: Aiming at the problem that it is difficult for a single acoustic feature to effectively represent language information in a noisy environment, a language identification method is proposed by combining mel-scale frequency cepstral coefficients and gammatone frequency cepstral coefficients. Firstly, the mel-scale frequency cepstral coefficients and gammatone frequency cepstral coefficients of speech are extracted. Then, the two features are transformed by matrix space to obtain the mel-scale gammatone cepstral coefficients of fusion feature. Finally, the fusion feature is input into the deep bottleneck network, and the language identification performance of MGCC features is tested in 25 different noise environments. The experimental results show that the identification accuracy of the proposed method is much higher than that of the single acoustic feature and other fusion features under different noise and different signal noise ratios. The accuracy of language identification can reach 99.56% in the clean corpus, and can still reach more than 93% under -5dB signal noise ratio, which proves the effectiveness and robustness of the proposed method.

Key words: language identification, fusion feature, deep bottleneck network, low signal noise rate, robustness