Cross spectral measurement of head related speech transfer functions using speaker’s own voice

Abstract

A cross spectrum method is applied to measure sound pressure variations around the head using the speaker’s own speech sounds. The variations are represented as transfer functions from the mouth reference point to a set of measuring points. Preliminary tests indicated that there are systematic frequency response variations depending on vowel colors. This vowel color dependency was not replicated in the classical measurement of speech radiation characteristics by J. L. Flanagan. However, taking into account the large (sometimes exceeding 20 dB) amount of variations, it is not likely to be negligible. A set of calibration and normalization procedures were introduced to reduce artifacts due to background noise, room acoustics, zeros in the speech spectra. A series of M-sequence based transfer function measurements were also conducted using a head and torso simulator to evaluate intrinsic errors in the cross spectral measurements. It was found that the standard errors in the cross spectral measurements using recorded speech sounds are around 1 dB. Based on these reference data and confidence interval calculations based on coherence, it is safe to conclude that the vowel color dependency is significantly modifying the transfer functions. [Work supported by JSPS.]