Improved spoken term detection using support vector machines based on lattice context consistency.