Speaker normalization for speech recognition using model-based vocal-tract parameters