Investigation of fixed-dimensional speech representations for real-time speech emotion recognition system

The real-time speech emotion recognition system is not only required to achieve the high accuracy, but also is needed to consider the memory requirement and running time in the practical application. This paper focuses on exploring the effective features with lower memory requirement and running time for the real-time speech emotion recognition system. To this end, the fixed-dimensional speech representations are considered because of its lower memory requirement and less computation cost. This paper investigates two types of fixed-dimensional speech representations which are high level descriptors and i-vectors and compares them with the conventional frame-based features low level descriptors in terms of accuracy and computation cost. Experimental results on IEMOCAP database show that although high level descriptors and i-vectors only contain the compact information comparing with low level descriptors, they achieve slightly better performance than low level descriptors. Experiments also demonstrate that the computation cost of i-vectors is much less than that of low level descriptors and high level descriptors.