Towards a classifier to recognize emotions using voice to improve recommendations

Abstract

The recognition of emotions in tone voice is currently a tool with a high potential when it comes to making recommendations, since it allows to personalize recommendations using the mood of the users as information. However, recognizing emotions using tone of voice is a complex task since it is necessary to pre-process the signal and subsequently recognize the emotion. Most of the current proposals use recurrent networks based on sequences with a temporal relationship. The disadvantage of these networks is that they have a high runtime, which makes it difficult to use in real-time applications. On the other hand, when defining this type of classifier, culture and language must be taken into account, since the tone of voice for the same emotion can vary depending on these cultural factors. In this work we propose a culturally adapted model for recognizing emotions from the voice tone using convolutional neural networks. This type of network has a relatively short execution time allowing its use in real time applications. The results we have obtained improve the current state of the art, reaching 93.6% success over the validation set.