Autocorrelation Analysis of Speech Sounds

A method of speech analysis using the short‐time autocorrelation function has been investigated. The conventional autocorrelation function, commonly used in studies of stationary time series, is here modified by including only a finite, weighted interval of integration. This modified result is termed the “short‐time” autocorrelation function. It is defined mathematically by the equation, φ11(τ, t) =  ∫ −∞∞ h(t − σ)f1(σ)f1(σ − τ)dσ, where f1(t) is the original time function, h(t) is a weighting function, and τ represents the delay time. When f1(t) represents a speech signal, φ11(τ, t) may be visualized as a three‐dimensional representation of the signal, similar to the well‐known intensity‐frequency‐time display. The two representations are related, since the Fourier transform of φ11(τ), the autocorrelation function, is the power spectrum. Correlation functions for a variety of speech sounds, spoken separately and in words, have been measured. Apparatus for accomplishing this includes an acoustic delay tub...