A statistical basis for objective measurement of speech levels

First-order probability distributions of speech amplitudes are studied to establish a theoretical basis for obtaining a measure of speech level. The logarithm of the long-term waveform of the speech envelope is found to be approximately uniformly distributed above a threshold. The average peak level (apl) is obtained by taking the time average of the log of the envelope waveform and deriving from it the peak of the log-uniform distribution which would have produced the same average. A theoretical analysis of various properties of the apl indicates that, within certain bounds, the apl satisfies a postulated set of requirements of an “ideal” speech level measure. A critical requirement is that the measure remain independent of the value of a threshold employed by a speech detector in the measuring device. It appears that variation in the threshold can typically change the apl by about one db. The Digital Speech Level Meter is described as an instrumentation of the technique used to obtain the apl. Measurements made with this meter are easily obtained and very repeatable, and are in general agreement with theoretical predictions.