Post-processing speech recordings during MRI

Abstract

We discuss post-processing of speech samples that have been recorded simultaneously during Magnetic Resonance Imaging (MRI) of the upper airways. Speech recordings contain acoustic noise from the MRI scanner. The required noise reduction is based on adaptive comb filtering designed for accurate formant extraction. Two kinds of speech materials were used to validate the post-processing algorithm. The primary material consists of samples of prolonged vowel productions during MRI. The comparison data was obtained from the same test subject, and it was recorded in anechoic chamber in a similar configuration as used during the MRI. Spectral envelopes and vowel formants were computed from the post-processed speech and from the comparison data. Vowel samples (with a known formant structure) were artificially contaminated using MRI scanner noise to determine performance of the post-processing algorithm. Resonances computed from a numerical acoustic model and spectra measured from 3D printed vocal tract physical models were used as comparison data. The properties of the recording instrumentation or the post-processing algorithm do not explain the observed frequency dependent discrepancy between the vowel formant data from two kinds of experiments: recordings during MRI and comparison data. It is shown that the discrepancy is statistically significant, in particular, where it is largest at ca. 1 kHz and 2 kHz. Numerical and experimental evidence suggests that the surfaces of the MRI head coil change the acoustics of speech which results in “exterior formants” at these frequencies. The discrepancy is too large to be neglected if the recordings during MRI are to be used for parameter estimation or validation of a numerical speech model, based on the MR images. However, the role of test subject adaptation to noise and constrained space acoustics during an MRI examination cannot be ruled out.