The paper presents a maximum-likelihood consistent bounded distance decoding algorithm for convolutional codes. The algorithm correctly decodes all error sequences which fall within the error-correcting sphere. A class of codes is defined, in which the decoder exploits the fact that only certain error sequences need to be corrected. For these codes the decoding is based on a reduced encoder state diagram. Thus, only a subset of the trellis or tree has to be searched in order to find the error pattern. An exact characteristication of the reduced state diagram is given in the paper along with an example.