Effect of fixed point computations on anger classification in speech signals

This paper investigates the effect of fixed point calculations on the accuracy of automatic emotion detection from speech signals. The tests used natural emotional speech recordings representing 16 speakers expressing two emotions: anger and neutral (unemotional) state. The feature set was derived from the Teager energy operator (TEO) and the speech was classified using the Gaussian mixture model (GMM) method. The results showed that with decreasing fixed point resolution from 16 bits down to 6 bits, the average classification error for the TEO increases from 0.0% to 6.6%. At 8 bit resolution, the error was an acceptable 2.4%, which implies that the TEO can be efficiently calculated using low-cost hardware.