Solving classification problems by Liquid State Machines (LSM) usually ignores the influence of the liquid state representation on performance, leaving the role to the reader circuit. In most studies, the decoding of the internally generated neural states is performed on spike rate-based vector representations. This approach occults the interspike timing, a central aspect of biological neural coding, with potentially detrimental consequences on the LSM performance. In this work, we propose a model of liquid state representation that builds the feature vectors from temporal information extracted from the spike trains, hence using spike synchrony instead of rate. Using pairs of Poisson-distributed spike trains in noisy conditions, we show that such model outperforms a rate-only model in distinguishing two spike trains regardless of the sampling frequency of the liquid states or the noise level. In the same vein, we suggest a synchrony-based measure of the separation property (SP), a core feature of LSMs regarding classification performance, for a more robust and biologically plausible interpretation.