Age identification from voice using feed-forward deep neural networks

In this paper, we propose to use feed-forward deep neural networks (DNN) for age identification from voice. We train two separate DNN using long-term and short-term features. The long-term features consist of various statistics of well-known short-term descriptors. We use mel-frequency cepstral coefficients (MFCC) as the short-term features. First a Gaussian mixture model (GMM) is trained using MFCC features, and then the GMM means are concatenated to obtain a GMM super-vector. The super-vectors are fed into the DNN. In the experiments, it is observed that the DNN yields very good recognition accuracy for age identification. Additionally, we observe that the age identification performance with the short-term features is better than the one with the long-term features.