Exploring Emotion Detection as a Possible Aid in Speaker Authentication

Voice as an authentication method is yet to be widely implemented as it is not yet as accurate as other authentication methods. In addition, it may be easy to coerce an authentic individual to authenticate themselves by means of voice. If an individual were authenticated under duress, the proper steps would need to be taken to mitigate any possible damage. Thus, this study attempted to verify the possibility of the addition of emotion detection as an aid in speaker authentication. Multiple experiments were conducted using two classifiers, namely a multilayer perceptron and a random forest. These experiments utilized different combinations of the mel-frequency cepstral coefficients, chroma, mel spectrogram, contrast, and tonnetz features. The experiments did not achieve the accuracy of other biometric authentication systems. However, they provided insight into the possible implementation of an emotion detection system and the value of each feature. Although the results from such a system may not be accurate enough to base an authentication decision on, additional security-related measures may be warranted if possible duress is detected.