Deep neural network combined posteriors for speakers' age and gender classification

Various applications depend on speakers' age and gender to satisfy their customer needs. The age and gender information of a speaker are concealed in the speech signal, which is vary in time duration and is exposed to background noise. These effects make the speakers' age and gender prediction one of the most challenging problems in the field of speech processing. Recently, remarkable developments have been achieved in the neural network field. DNN is considered as one of the state-of-art classifiers. Due to the success of DNN in many speech applications, we propose combined posteriors of the softmax and last hidden layer for class model representation based on a trained DNN. DNN is used in this work to extract the softmax and last hidden layer posteriors from the concatenated frames for each utterance. The outputs of the combined posteriors for each frame in the training database are accumulated to form a new model for each class. Our experimental results showed that the proposed class model is capable of representing speakers age and gender for new utterances. The performance of the proposed work is evaluated by using GMM-UBM model on a publicly available Age-Annotated Database of a German Telephone Speech Database (aGender). The proposed work outperforms the baseline system by approximately 6% in terms of overall accuracy.