Sparse kernel-based models for speech recognition door