This paper describes three different approaches to the use of spectral transformation for supervised speaker adaptation in continuous speech recognition. Each approach may involve transforming feature vectors of the speech of a speaker, transforming mean vectors of the HMMs of a reference recognition system, or transforming both feature vectors and HMMs. A comparison of these approaches is investigated using the ARPA 1000-word Resource Management (RM1) continuous speech corpus. Using the average speaker-independent (SI) test result as a reference point, it is found that the best adaptation approach can achieve an error reduction of 23.5% by using 10 sentences as adaptation speech.