This paper deals with joint gender identification and age group classification from speech, aimed at improving the functionalities of Interactive Voice Response Systems. Deep Neural Networks are used, because they have recently demonstrated discriminative and representation capabilities over a wide range of applications, among them, speech processing problems based on features extraction and selection. A comparative study of various neural network architectures and sizes is presented to gather knowledge about performance dependence on the network architecture and the number of free parameters. The classification framework was trained and evaluated using Mozilla's 'Common Voice' dataset, an open and crowdsourced speech corpus. The results are promising, with the best systems achieving a gender identification error lower than 2% and an age group classification error lower than 20%.
The goal of the system presented in this paper is to develop a natural talking gesture generation behavior for a humanoid robot. With that aim, human talking gestures are recorded by a human pose detector and the motion data captured is afterwards used to feed a Generative Adversarial Network (GAN). The motion capture system is capable of properly estimating the limbs/joints involved in human expressive talking behavior without any kind of wearable. Tested in a Pepper robot, the developed system is able to generate natural gestures without becoming repetitive in large talking periods. The approach is compared with a previous work, in order to evaluate the improvements introduced by a computationally more demanding approach. This comparison is made by calculating the end effectors’ trajectories in terms of jerk and path lengths. Results show that the described system is able to learn natural gestures just by observation.
