A speech-to-noise ratio measurement algorithm.

Abstract

An algorithm to measure speech-to-noise ratios has been implemented on a minicomputer. The algorithm attributes the energy within each consecutive 20-ms frame of a speech-plus-noise waveform to either a speech or noise source. This discrimination process is based upon the known characteristics of frame energy histograms of such waveforms. In response to observed inaccuracies of this discrimination process in cases of low speech versus noise separation, a method of estimating the speech Vrms of the signal is incorporated which attempts to recover speech energy, "masked" by noise. The algorithm's ability to track known speech-to-noise ratios on a decibel-for-decibel basis down to a ratio of approximately 5 dB has been demonstrated by experimentation.