Sparse kernel-based models for speech recognition by