Key points are not available for this paper at this time.
The phoneme identification process of an automatic speech recognition system may be aided through the use of statistics of phoneme occurrence in conversational English. These statistics are also applicable to the fields of linguistics and speech, to teaching English as a foreign language and to speech pathology. In this study a data base containing 103,887 phoneme occurrences taken from casual conversational American English was obtained through interviews of sixteen adult males and ten adult females. The speech was transcribed using a quasi-phonemic system, known as ARPAbet, plus selected phoneme alternates and was analysed with computer assistance to obtain the rank order of phonemes according to frequency of occurrence. Also, the radius of the confidence interval for the observed frequency of occurrence was calculated at the 95% level for each phoneme. The top ten phonemes (in order, / a, n, t, i, s, r, i, l, d, ε /) account for 47% of all the data. As expected, the results of the present study correlate highly with those of one other major study of natural speech. Comparisons show some interesting differences in detail, however, that appear to be attributable to relatively minor variations in the experimental procedures.
Mines et al. (1978) studied this question.