The most curious thing about the program is that, to identify the mood of the person speaking, he does not look atwhatsays but inasit says. “In fact we are using recordings of actors reading dates on the calendar, what interests us is not the content but the way of saying it,” explains Wendi Heinzelman, one of the creators of the algorithm. “For example, you hear someone speak and you think they are angry,” explains Melissa Sturge-Apple, a psychologist who helps on the job. “But what makes you think that?Emotions affect the way we speak and alter the volume, tone, and harmony of speech“, he adds. The program analyzes twelve characteristics of speech, such as tone and volume, to identify one of the five basic emotions chosen: sadness, happiness, fear, anger or neutrality. The results achieve a81 percent accuratein the prognosis, a figure greatly improved compared to the 55 percent achieved in other studies.
The first prototype of the application shows a happy face or a sad face based on the results of the speech analysis. “The research is just beginning,” explains Heinzelman, “but it is easy to imagine a more complex application that could, for example,adjust screen colors or play musicadapted to the mood of the mobile owner “.
The team is now working on improving the precision of the estimates, since these are good if the device is always ‘trained’ with the same voice, but when different voices are used, the system does not recognize the interlocutor’s mood as easily.
