High-level approaches to confidence estimation in speech recognition