Fostering Positive Team Behaviors in Human-Machine Teams through Emotion Processing: Adapting to the Operator's State
Technical Report,22 Feb 2017,19 Jun 2018
Royal Melbourne Institute of Technology Melbourne Australia
Pagination or Media Count:
The team developed a software system that simultaneously recognizes 7 emotional categories as speech is produced. It is suitable for applications on cellular phones and online speech communication platforms. The methodology uses deep learning DL with speech signals being represented in the form of RGB images of speech spectrograms. By representing speech signals in the form of RGB images, the speech classification problem was re-defined as an image classification task. This created an opportunity to replace the lengthy and data-costly training of a deep neural network by the shortened and more data-efficient fine tuning of an existing pre-trained image classification network AlexNet. The speech emotion recognition SER results achieved with the finetuned AlexNet FTAlexNet showed an average accuracy of 80 for the Berlin Emotional Speech data. This result was found to be comparable with existing state-of-the art techniques, but with the advantage of significantly lower computational and data costs. The ability to analyze emotions represented in speech was then applied to a multi-stage classification system with intermediate learning MSIL. In this scheme, the system can leverage the mistakes made by the primary stage in learning from the data and use them to improve the learning by the secondary stage. It is felt that this approach can be incorporated as a building block for more complex multi-level machine reasoning systems.