Statistical properties of spectral samples derived from the continuous speech of six talkers and summarized by means of covariance‐matrix eigenvectors are used to study the dimensionality of the data space. The importance of eliminating low‐level samples by means of a fixed threshold is emphasized, and criteria for selecting such a threshold are presented. Measurements of spectral correlations stabilize after about 30 sec of speech, suggesting that short‐term examination of a talker's output may prove sufficient to calculate parameters useful in recognition schemes. Some features of the correlation matrix, which are readily displayed via isocorrelation contours, appear to be related to talker characteristics, while others are talker independent. The results suggest that the separation of speech data into gross classes prior to the application of statistical procedures will enhance the performance of processing schemes.