A number of current approaches to automatic speech recognition make use of a preliminary analysis of the incoming speech signal into discrete phonetic units. To do this, two types of decision have to be made: one is the identification of boundaries between neighbouring segments, and the other is the identification of a segment as belonging to a particular broad phonetic category. Little has been written about ways of assessing the degree to which a particular machine-generated phonetic transcription matches the transcription produced for the same data by an expert human; this paper discusses some of the problems involved.