This paper presents a voice conversion algorithm based on Hidden Markov Models that does not requires explicit phonetic labeling of the input speech. Additionally, the proposed voice conversion algorithm also uses an excitation estimation algorithm previously presented by the authors to achieve higher speech quality without compromising speaker identity conversion. The performance of the proposed algorithm was compared, using listening tests, with the performance of a recent voice conversion algorithm based on HMM but requiring phonetic labeling. The proposed algorithm was found to achieve equivalent identity conversion scores while improving the perceived quality of the converted speech. Thus, the proposed algorithm was found as a viable alternative for conversion applications where phonetic labeling is not practical.