Adaptive measurement of crossover frequencies for intelligibility prediction

Abstract

In SII theory, frequencies where speech spectra can be divided into two equally-intelligible subbands are called crossover frequencies. These frequencies play a crucial role in SII calculations, and also designate spectral regions that contain important speech recognition cues. Typically, crossover frequencies are found by measuring psychometric curves for speech processed by a series of low-pass and high-pass filters, and then finding the two curves’ intersection: an inefficient, time-consuming process. The present study introduces an up/down quantile estimation algorithm that adaptively steers filter cutoff frequencies toward the crossover frequency. Changes in cutoff frequency are governed by comparisons of block trials for low-pass and high-pass filtered speech that meet theoretical requirements for convergence toward the crossover frequency. Preliminary results for trials with nonsense syllables show that the proposed method’s estimates match those obtained in published trials using the conventional method. Applications in SII measurements and speech recognition cue measurement will be discussed.