Spectral Discontinuity in Concatenative Speech Synthesis - Perception, Join Costs and Feature Transformations