GAN-Based Augmentation for Gender Classification from Speech Spectrograms

The focus of this study is on gender classification from speech signals produced by adults. Automatic estimation of gender has a broad variety of applications ranging from forensics, authentication systems, diarization of meetings, or user-centered interactive agents, and plays a crucial role also in ‘internal’ technological solutions aimed at improving model accuracy, such as selection of gender-specific acoustic models. In this study, we explore scenarios where only a limited amount of real training data is made available for training of a 2-D convolutional neural network classifier. To address sparsity of the training data, the training set is augmented by synthetically generated samples produced by a generative adversarial network. The adversarial network is given access only to the same limited real-world training dataset as the gender classifier. We demonstrate that even when 80% of the already limited training data are removed and replaced by synthetic spectrograms, the gender classifier models can still be successfully trained thanks to the augmentation and maintain competitive performance.