Computational model for the modulation of speech-in-noise comprehension through transcranial electrical stimulation

Background — Transcranial electrical stimulation (TES) can non-invasively modulate neuronal activity in humans. Recent studies have shown that TES with an alternating current that follows the envelope of a speech signal can modulate the comprehension of this voice in background noise (Wilsch et al., 2018). However, how exactly TES influences cortical activity and influences speech comprehension remains poorly understood. Here we present a computational model for speech coding in a spiking neural network and employ it to investigate the effects of TES on the coding of speech in noise.

Methods — Based on previous work, we established a computational model of a spiking neuronal network that encodes natural speech through entraining network oscillations in the theta and gamma frequency range (Hyafil et. al., 2015). We used the network’s spiking output to classify speech in different levels of background babble noise. We then investigated the effect of different external currents on the network dynamics as well as on the neural output and the associated speech coding. Finally, we analysed the behaviour of the computational model and its speech classification performance in different conditions to optimize the stimulation paradigm for enhancement of natural speech processing.

Results — The computational model generated coupled oscillations in the theta and the gamma frequency range. In agreement with experimental results, the slower theta oscillations reliably predicted the onsets of syllables and provided a temporal reference frame for the faster activity in the gamma band that encoded phonemes. Classifying speech in different levels of background noise yielded results comparable to normal human performance, with a 50% speech recognition threshold at approximately -1 dB SNR. Simulating the effect of simultaneous external current with a range of different temporal patterns and stimulation intensities we were able to identify the parameters that impeded as well as enhanced the neural coding of speech in noise.

Conclusions — The developed model provides an insight into the neural mechanisms through which speech in noise can be processed in the auditory cortex and how TES can enhance this processing. Moreover, our computational model allows to optimize the temporal pattern of the stimulation for improving speech-in-noise comprehension.

References
Hyafil et al. (2015) Elife, 4, e06213.
Wilsch et al. (2018) NeuroImage, 172, 766-774.

Acknowledgements
This study was supported by the EPSRC Centre for Doctoral Training in Neurotechnology for Life and Health. We thank Alexandre Hyafil, Shabnam Kadir and Milos Cernak for helpful comments and fruitful discussions.