Direct and joint-space approaches to the use of spectral transformation for speaker adaptation in continuous speech recognition.