Music and Speech Elicit Similar Subcortical Responses in Human Listeners

Abstract

Music and speech are two sounds that are unique to human beings and encountered in daily life. Both are transformed by the auditory pathway from an initial acoustical encoding to higher level cognition. Most studies of speech and music processing are focused on the cortex, and the subcortical response to natural, polyphonic music is essentially unstudied. The goal of this study was to compare the subcortical encoding of music and speech using the auditory brainstem response (ABR). While several methods have recently been developed to derive the ABR to continuous speech, they are either not applicable to music or give poor results. In this study, we explored deriving the ABR through deconvolution using three regressors: 1) the half-wave rectified stimulus waveform, 2) the modeled inner hair cell potential, and 3) the auditory nerve model firing rate (ANM), where the latter two were generated from a computational auditory periphery model. After quantitatively comparing the three regressors, we found the ANM regressor yields robust and interpretable ABR waveforms to diverse genres of music and multiple types of speech stimuli. We then used the ANM-derived ABRs to compare the subcortical responses to music and speech and found that they are highly similar in morphology. We further investigated cortical responses using the same deconvolution method, and found the responses there were also quite similar, which was unexpected based on previous studies. Our study first developed approach that can be used to investigate subcortical encoding and processing for both natural speech and music with the potential for use in cortical encoding. Based on that, we concluded that when accounting for acoustical differences' nonlinear effects on peripheral encoding, the derived music and speech brainstem responses are highly correlated.

Predictors for estimating subcortical EEG responses to continuous speech

“…It must be noted that although the wave V peak was used as the primary metric of performance, the conventional click ABR consists of several other morphological features [36]. The wave V peak was selected here to both be consistent with prior work [6, 19, 16], and because it was the only consistent feature that was detected in all subjects. TRFs using ZIL had shorter wave V peak latencies (see Figures 1 and 3).…”

“…Since brainstem responses are largely agnostic to stimulus polarities, a pair of predictors were generated for each model, using an input stimulus pair with the original stimulus and the stimulus with opposite sign. In line with prior work [6, 19], TRFs were fit to each polarity separately and then averaged together.…”

“…For all models, the input was the audio stimulus, as recorded by the Strim-Trak system. The lags inherent in the output of each model were accounted for by shifting the generated predictors to maximize the correlation with the rectified speech predictor, similar to prior work [19]. Since brainstem responses are largely agnostic to stimulus polarities, a pair of predictors were generated for each model, using an input stimulus pair with the original stimulus and the stimulus with opposite sign.…”

“…The implementation in the Python cochlea package [28] was used with 43 auditory nerve fibers with high spontaneous firing rates and center frequencies logarithmically spaced between 125 Hz and 16 kHz, in line with previous work [19]. To speed up computation, an approximation of the power-law adaptation was used [19]. The outputs of this model are the mean firing rates of the auditory nerves, which were averaged to form the final predictor pair.…”

“…Another concern is that TRFs which linearly map the sound stimulus to the EEG ignore several highly non-linear processing stages in the auditory periphery [18]. One study has recently shown that predictors derived from auditory models that incorporate non-linear stages can lead to improved subcortical TRFs [19]. Another recent study showed that auditory-model-derived predictors outperform previously used envelope predictors even for cortical TRFs [20].…”