This study investigated the neural representations of speech sounds in the auditory nerve, inferior colliculus, and auditory cortex of guinea pigs using degraded speech stimuli. Neural responses were recorded from these three brain regions. The study found that representations in the auditory nerve and inferior colliculus remained highly discriminable even with severe degradations to the speech sounds and longer smoothing of the neural responses, unlike representations in auditory cortex which required optimal smoothing. Representations were also more redundant in cortex compared to lower regions for speech discrimination tasks. The discriminability of peripheral representations was more robust to spectral degradations than what is observed in human perception.

1. 3. Natural speech
2. Methods
1. Introduc�on
Inves�ga�ng the neural code for speech discrimina�on
MRC Ins�tute of Hearing Research, No�ngham, UK
mark.steadman@ihr.mrc.ac.uk
Mark Steadman & Chris Sumner
4. Vocoded speech
5. Summary
Speech recogni�on is robust to severe degrada�ons e.g. cochlear
impants, vocoded speech.
Human psychophysical studies using degraded speech sounds cons�tute
a powerful toolset to explore the neural bases of speech percep�on.
Inves�ga�ng which aspects of the neural representa�on facilitate robust
speech discrimina�on, and how this changes throughout the auditory
brain, increases our understanding of neural processing of complex
signals.
Medial consonants elicit dis�nct pa�erns of ac�vity in the auditory nerve, inferior
colliculus and auditory cortex of anaesthe�sed guinea pigs.
The efficacy of spike �ming depends on the nature of the task (speech token vs.
consonant recogni�on) and the brain region.
Unlike percep�on, discriminability of peripheral representa�ons is highly robust to
severe spectral degrada�ons in the s�mulus, provided temporal cues are preserved.
Cor�cal representa�ons of speech sounds are less redundant for a speech
discrimina�on task than those in the more peripheral nuclei.
Shannon, Robert V., et al. "Speech recogni�on with primarily temporal cues." Science 270.5234 (1995): 303-304.
References:
Computa�onal model (auditory nerve)
Outer/middle
ear filter
Nonlinear basilar
membrane filterbank
Inner hair cell
transduc�on model
Stochas�c spike
genera�on
Neural classifica�on
Neural
Classifier
% Correct
�me since onset, ms
neuron#
300 400 500
/asa/ neurogram
(auditory nerve)
Auditory Nerve Auditory CortexInferior Colliculus
200 300 400 200 300 400200 300 400
/ada/
200 300 400 200 300 400200 300 400
/asa/
200 300 400 200 300 400200 300 400
/ama/
Figure: Example neurograms from each brain region for 4 of the 16 consonants.
1 10 100100 101 102
0
20
40
60
80
100
100 101 102
percentcorrect
one talker three talkers
Auditory nerve Inferior colliculus Auditory cortex
smoothing window length, ms number of recording sites
chance
100 sites 106 sites 196 sites
Figure: Discriminability of representa�ons of vocoded speech with op�mal smoothing. Human psychophysics
taken from Shannon et al. (1995).
Electrophysiology
Inferior colliculus
Auditory cortex
in vivo extracellular recordings in the inferior
colliculus and auditory cortex of anaesthe�sed
guinea pigs.
S�mulus set parametrically degraded using a noise vocoder using 1, 2, 4 and 8
spectral channels (between 0.1 and 4 kHz), with 16 and 500 Hz envelope rates.
S�muli
16 a-consonant-a sequences.
3 male talkers.
Parametrically degraded using a "noise vocoder".
Bandpass
filter
Envelope
extraction
Noise
Bandpass
filter
0 100 200 300 400 500 600
0
1
2
3
4
original s�mulus
�me, ms
frequency,kHz
0 100 200 300 400 500 600
0
1
2
3
4
vocoded s�mulus
�me, ms
frequency,kHz
Mul�-unit spiking ac�vity extracted by
filtering and thresholding raw signal.
Nearest-neighbour based classifier trained on
average "neurograms".
Average neurograms were (op�onally) combined
across talkers.
Neurograms are allowed to temporally shi�
rela�ve to each other such that distances were
minimised.
Neurograms were parametrically smoothed by row-wise convolu�on
prior to classifier training and tes�ng.
For a single talker speech token discrimina�on task, representa�ons in the auditory
nerve and IC are prefectly discriminable regardless of smoothing window length.
Unlike data from an analogous human psychophyics task, high rate envelope cues
increase the discriminability of neural responses In the auditory nerve and IC This is
not the case in the auditory cortex.
This benefit is only apparent where millisecond-precise spike �ming is considered in
the auditory nerve, but is s�ll apparent in the IC where longer smoothing windows
are used.
Cor�cal representa�ons are perfectly discriminable with op�mal smoothing windows.
Where the classifier was trained using representa�ons combined across talkers,
discriminabilty is highly depended on choice of smoothing window.
For a comparable number of neurons, representa�ons are far more discriminable in
the auditory nerve and inferior colliculus than in the cortex.
Decreasing the spectral complexity of the signal has less effect on the discriminability
of peripheral representa�ons than it has on percep�on.