Unsupervised Personalization of an Emotion Recognition System: The Unique Properties of the Externalization of Valence in Speech