Methods of Measuring Speech Spectra
S. S. Stevens, James P. Egan, George Armitage Miller · The Journal of the Acoustical Society of America · 1947
A comparison was made of several alternative methods and devices for analyzing speech in terms of its acoustic spectrum (a time average of the sound-pressure level per cycle vs. frequency). A general procedure, applicable to a wide variety of analyzing instruments, consists of a comparison made between a known acoustic spectrum and an unknown spectrum (speech). For maximum convenience, the known spectrum should be a “white noise” whose spectrum level is a constant number of decibels per cycle. This spectrum is used as a standard to provide an over-all calibration of the entire recording and analyzing system. In practice, the shape of the speech spectrum is obtained directly as the difference in decibels, in each pass band of the analyzer, between the level obtained for the white noise and the level obtained for the speech. In order to measure conveniently the spectrum of the noise used as the standard of reference, a graphic analyzer was devised as follows: A Hewlett-Packard wave analyzer, Type 300A, was modified to give access to the 20,000-cycle voltage al the input to the meter system. This voltage was used to drive either of two types of graphic recorders. Continuous analyses from 0 to 16,000 c.p.s. can be made with various constant band widths. The procedure of comparing a known with an unknown spectrum eliminates the necessity of calibrating separately the recording, reproducing, and analyzing equipment. In the analysis of speech, only the calibration of the microphone need be considered as a final correction to the data. This procedure also makes it possible to obtain accurate analyses despite the use of voice recording and reproducing equipment having unknown frequency characteristics. Although the choice of filter cut-off frequencies is arbitrary, it has been found most convenient to analyze speech by dividing it up into bands that stimulate equally wide regions on the basilar membrane. This is accomplished by choosing filter cut-offs at equal intervals along the mel scale of subjective pitch. Within the limits of observational error the same answer is obtained when speech is analyzed by any of three different measuring systems: (a) square-law integrator (audio spectrometer), (b) linear integrator, (c) R.C.A. noise meter and Esterline Angus graphic recorder. When, with any of these systems, a speech spectrum is determined as the difference in decibels between the two analyses — that for white noise and that for speech — the speech spectrum can be stated in terms of sound-pressure level per cycle averaged over each of the nominal pass bands of the filters. Very narrow pass bands (5 c.p.s.) reveal details in the speech spectrum not disclosed by the wider filter bands.