I tried to estimate total variable space (T) for i-vector estimation, in speaker recognition, but if i select the factor number to more than 100 it gave wrong results but with 100 or less it provided good result. So how to decide that we need as much number of factor (t dimension) for total variablity space?