Speech recognition is carried out by matching parameterized speech with a dynamically extended network of paths comprising model linguistic elements (12b, 12c). The units are context related, e.g. triphones. Some elements cannot be converted to models at the time when it is necessary to incorporate the element into the paths because the context is not defined at the relevant time. In order to allow transfer the element is transferred as a place marker (21, 22, 23 24) which is converted when a later extension completes the definition of the triphone. The place markers (12a) can be used to identify the locations for subsequent extensions.