Flexible speaker adaptation for large vocabulary speech recognition