Use of a Correlational Method to Estimate a Listener's Weighting Function for Speech

The purpose of this study was to determine if it is feasible to use the correlational method (Lutfi, 1995; Richards and Zhu, 1994) to estimate how listeners use or weight the information contained within various frequency bands of speech. Three naturally spoken vowel-consonant-vowel (VCV) syllables (/aba/, /aga/, and /ada/) were presented monaurally to listeners. Each of the VCV waveforms were filtered into three separate frequency bands (i.e., low, mid, and high). Each band was then independently and randomly degraded at various signal-to-noise (S/N) levels (-7, -5, -3, -1, or +1). On each trial, listeners were asked to identify the VCV that was presented to them. For each trial, the S/N level of each frequency band, the stimulus that was presented, and the listener's responses were all recorded and stored in a file. From this trial-by-trial data, a point biserial correlation was computed between the listener's response (correct or incorrect identification) and the degradation within each frequency band. The stronger the correlation, the greater influence that given frequency band had on the listener's performance on the task. From these relations it was shown that it is possible to obtain a listener's weighting function for speech. Results showed that although most listeners weighted the mid-frequency band the greatest, several of the listeners used different weighting strategies to perform the task. Several methodological issues are discussed in regard to improving the future application of the correlational method to speech.