In recent studies involving the correlation of speech segment durations, the measurement error in locating segment boundaries has usually been ignored or assumed to be constant and small. Since the relative magnitude of the error in locating such a boundary may vary greatly depending upon its phonetic context and the procedures employed, thereby affecting the derived intersegment covariances profoundly, an experiment was designed to estimate these errors directly. Several tokens of four sentences containing syllables differing in the quality, length, degree of stress, and consonantal context of their nuclear vowels were copied onto two‐track tapes and sent to several professional phoneticians with instructions to measure the durations of the indicated vowels, on first one track and then, at least two weeks later, the other. The resulting measurement errors varied systematically in several ways: (1) long, stressed vowels had greater errors than short, unstressed ones; (2) initial voiced stops (/b/) were easiest to segment and approximants (/w/) hardest, with voiceless stops (/k/) and fricatives (/f/) intermediate; (3) great differences among phoneticians did not appear to depend in any obvious way upon the instrumentation used. The minimum standard deviation for measurements of vowel duration by the most accurate phoneticians ranged from about 2 to 5 ms; 95% confidence intervals for carefully measured vowel durations may therefore be expected to range from about 10 to 25 ms, depending upon the consonantal context of the vowel. Because of this range of intrinsic measurability of vowel durations, and therefore presumably other phonetic segments as well, the results of studies using covariance statistics of segment durations should be interpreted cautiously.