The paper describes the development of a non-linear temporal compression algorithm applied to a sequence of phonetic classification probability vectors. This is employed as a means of data reduction between the two stages of a recogniser performing lexical access based on recurrent neural nets. The algorithm operates by identifying pseudo-stationary segments in the sequence of probability vectors, and forming the average vector of each segment thus identified. Boundaries between pseudo-stationary segments are located using a metric applied to adjacent vectors. Various metrics are compared, along with two heuristics and a novel NN technique which perform the segmentation task using the temporal profile generated by the metric. The compression ratios may be varied smoothly between 3.7 and 6.8, under the control of a threshold parameter with an associated segment deletion rate of between 3% and 27%.