An unsupervised speaker adaptation method is proposed for application to a speaker independent recognition system which uses a demi-syllable based, continuous mixture-density HMM. The spectral interpolation technique, which has been used in supervised adaptation, and a scheme of utilizing recognition outputs of recognition system are employed. In an experimental application employing results obtained from a 5000-word Japanese vocabulary set recognition task, the error rate was reduced with the method from 15.5 % to 8.7 %, which is comparable to the error rate obtained by supervised adaptation. The method has also shown promise in the utilization of results obtained from a connected syllable recognition task in which the Japanese vocabulary set was unlimited.