In the conventional hidden Markov model, the model parameters are reestimated by an iterative procedure known as the Baum-Welch method. This paper proposes an alternative procedure using fuzzy estimation, which is generalised from the fuzzy c-means and the Baum-Welch methods. An extension of this approach, which uses a garbage state to deal with outlier data is also proposed. Experiments using the TI46 speech data corpus show this approach can be applicable to speech and speaker recognition.