One method to communicate in a very high noise environment is to use bone-conducted speech. In this paper, a technique of improving the quality of bone-conducted speech is presented. The bone-conducted speech signal of a speaker is passed through a reconstruction filter designed by using the long-term spectra of the bone-conducted and normal speech signals. Properties of the normal, bone-conducted and reconstructed speech signals are investigated and it is shown that for speaker recognition, the reconstructed speech signal could be usefully utilized.