German researchers from the Max Planck Institute for Human Development in Berlin have discovered that some artificial intelligence (AI) models can identify human emotions from short audio recordings as well as humans. The article was published in the scientific journal magazine Frontiers in Psychology (FIP).
“We showed that machine learning can be used to recognize emotions in audio clips as short as 1.5 seconds. Our models achieved human-like accuracy in classifying meaningless, emotionally charged sentences spoken by actors,” said lead author Hannes Diemerling.
Researchers used gibberish in Canadian and German to see whether AI could accurately identify emotions regardless of their context and cultural-linguistic nuances.
Using training data, scientists have created artificial intelligence models that work in one of three ways. Deep neural networks (DNNs) are like complex filters that analyze components of the voice such as frequency or pitch (for example, the voice gets louder because the speaker is angry) to reveal hidden emotions.
Convolutional neural networks (CNNs) scan for patterns in the visual representation of movie soundtracks, similar to detecting emotions from the rhythm and texture of sound.
The hybrid model (C-DNN) combines both methods, using both auditory and visual spectrogram to predict emotions. The models were then tested for performance on both data sets.
“We found that DNN and C-DNN achieved better accuracy than using spectrograms alone in a CNN,” Diemerling said.
The results showed that it is possible to develop systems that can instantly interpret emotional signals and provide immediate and intuitive feedback in a wide range of situations.
This can help create scalable and cost-effective applications in fields as diverse as therapy and interpersonal communication technologies.
Previous scientists educated AI thinks through its answers and improves its efficiency.