Obstruent acoustic landmark enhancement for cochlear implants

Abstract

Cochlear implant users are typically able to achieve high levels of speech recognition in quiet
but struggle to follow speech even in moderate levels of background noise. It may be possible
to improve cochlear implant users’ speech perception scores in noise, by making more
efficient use of the limited bandwidth available for transmitting important speech information,
rather than increasing the amount of information transmitted. Acoustic landmarks are
locations in the speech signal which are rich in information and allow a listener to identify a
particular speech sound as a vowel, sonorant consonant or obstruent consonant; it is around
these regions that important speech cues tend to be concentrated. Obstruent consonants are
signalled by sudden amplitude and spectral changes and the onset/offset of a period of noise.
It has been shown that the auditory system is particularly responsive to rapid spectral changes,
manifested as increased firing rates of auditory nerve fibres, particularly at onsets of signals.
Cochlear implant users commonly confuse speech sounds with rapidly changing spectral
patterns, possibly due to the poor transmission of obstruent landmark information.
The aim of the present work was to develop an obstruent landmark enhancement strategy
which could be integrated into current cochlear implant processing. The first stage of this
process required the identification of obstruent landmarks from the noise-mixed speech
stimuli. An existing automatic landmark detection algorithm did not achieve the high levels
of accuracy required for use in the present study and so a set of hand-generated labels were
created, and used to guide the proposed obstruent landmark enhancement strategy. A series of
cochlear implant simulation experiments were conducted to help evaluate the strategy and
guide further developments. Results from the simulation studies suggest that the proposed
method of obstruent landmark enhancement does not help to improve speech recognition in
noise for normal hearing listeners listening to a cochlear implant simulation. It is likely that
the strategy outlined in this thesis did not help to improve the saliency of obstruent landmark
events as the enhancement was applied to the noise as well as the target speech signal,
making it difficult for listeners to resolve the boosted landmark information. However, the
results also highlight the limitations of using cochlear implant simulations to evaluate the
strategy and so the findings are not necessarily a predictor of actual cochlear implant user
performance.

More information

Identifiers

Catalogue record

Export record