Evaluation of a signal-averaging method for the measurement of signal-to-noise ratios in speech

Abstract

Many abnormal voice qualities are characterized by unusually large amounts of noise. Although the presence of noise is easy to identify, precise quantification of signal-to-noise (S/N) ratios is a difficult problem. Yumoto, Gould and Baer [J. Acoust. Soc. Am. 71, 1544–1550 (1982)] developed a technique based on signal averaging. Using a sustained vowel segmented into pitch periods, the signal component of the S/N calculation is defined as the average of the individual pitch periods. The noise component is estimated by successive subtractions of the average signal from individual pitch periods of the original vowel. In the present study, the accuracy of this technique was evaluated in a series of computer simulations. A synthesized vowel was mixed with varying amounts of synthesized noise. S/N ratios varied over a 36-dB range. Known S/N ratios were compared with S/N ratios calculated by the Yumoto et al. technique. The correlation between calculated and actual S/N ratios was nearly perfect (r = 0.99) and the mean absolute error was only 0.1 dB. Results are also reported for tests involving naturally produced vowels mixed with synthesized noise. [Work supported by NIH.]