Error Prediction in Spoken Dialog: From Signal-to-Noise Ratio to Semantic Confidence Scores