• No results found

Acoustic modelling of CI processing and electrical/neural interface

Chapter 2. Background

2.5. Acoustic models of CI function

2.5.3 Acoustic modelling of CI processing and electrical/neural interface

The main processing parameter which has received attention in AM studies is channel number. A number of the studies cited in 2.4.2 used AM results exclusively or

compared AM results with equivalent CI user data. To date there have been no studies of changes in CI temporal processing characteristics using AMs. One of the possible problems with such studies is the spectral distortion caused by changes in modulation rate with sine wave carriers, as noted in 2.5.1. It is unclear whether or not AMs using noise band or sine wave carriers can provide appropriate models of changes occurring to temporal information with higher stimulation rates. One of the aims of the work reported in chapter 5 was to evaluate changes in stimulation rate using an AM in parallel with CI users, in order to determine whether the model provides an accurate representation of the changes in temporal information (or the absence of such

changes) associated with stimulation rate. However, it should be made clear that such a model is only able to deal with one of two distinct aspects of stimulation rate. These two aspects are i) changes in neural response associated with increased number of pulses presented and ii) changes in temporal sampling associated with changes in stimulation rate. A model of the first aspect is beyond the scope of the present study, and would require a more sophisticated estimate of the physiological response and an understanding of what stimuli would be necessary to engender an analogous response in a NH listener. Rather, the purpose of using an AM in the context of stimulation rate is to replicate the changes in analysis filterbank output that occur with changes in stimulation rate. For CI systems which implement an IIR filterbank, analogous AM data have been obtained by looking at changes in envelope cut-off frequency which would be set at some value less than half the stimulation rate (see 2.3.3 for a brief discussion of some papers looking at variations in envelope cut-off frequency). However, with the Nucleus 24 system, the purpose of an AM is to represent the

changes in envelope bandwidth that occur as a function of changing overlap between FFT analyses, as this is the parameter which varies in accordance with stimulation rate changes (as described in 2.3.3). The AMs used in the experimental work reported in chapters 4 and 5 used modulated carrier stimuli at whatever rate was determined by the output of the analysis filterbank; therefore, if the analysis filterbank were able to convey increases in envelope modulations with increased FFT overlap (the parameter associated with increased stimulation rate), this increase in envelope bandwidth would be represented accordingly in the carrier stimuli. In practice, as noted in 2.3.3, the effective envelope bandwidth of the fixed FFT length Nucleus 24 processor appears to be limited by the maximum non-overlapping FFT analysis rate (125 Hz), irrespective of stimulation rate, and this suggests that variations in envelope bandwidth across stimulation rates are minimal. However, the AM provides a faithful reflection of the changes in filter output as a consequence of increased stimulation rate. A caveat should therefore perhaps apply that the term “stimulation rate”, as applied to the Nucleus 24 AM stimuli, really means “envelope bandwidth as a function of FFT analysis overlap concomitant with stimulation rate changes”. This does not mean that the AM can be a good model of neural changes occurring as a consequence of

stimulation rate which are independent of changes in temporal sampling (if these occur, this would be shown by increased performance in CI users with higher rates but not associated increase in performance with the AM).

It is worth noting some of the limitations of the evidence base from AM studies. The first relevant point is the majority of AM studies have used fixed-channel IIR filter processing (although Dorman et al. (2002) is an exception to this) and therefore cannot strictly be considered as appropriate models of signal processing using a peak- picking strategy such as ACE, or, in any case, of processors which implement a FFT filterbank. A more general limitation of AM studies to date is that those studies which have compared AM performance with CI user performance directly have used CI users with varied processing parameters and devices, making direct comparison with a specific set of processing parameters impossible. For example, Fu and Nogaki (2005) compared 10 CI subjects with 6 NH subjects listening to an AM. The CI users were a varied group: 4 were users of the Nucleus 22 device, one was a user of the MED-EL device, one a user of the Clarion 1 device, while 4 were users of the Clarion CII

channel number and stimulation rate all varied across CI subjects. The AM used was a noise band model with a frequency range of 200 to 7000 Hz and 16 channels. The lack of close correspondence between processing parameters used in an AM and those used in a matched group of CI users means that the importance of the specific

parameters is unclear.

As discussed in 2.4.2, one of the perceptual consequences of cochlear implantation is an effective upward shift in perceived frequency compared to NH. A few studies have attempted to incorporate these pitch shift characteristics of CI stimulation into an AM by using a mapping between analysis and carrier frequencies derived from Greenwood's work (Greenwood, 1990). Some studies have sought to compare AM performance with and without pitch-mismatch. Shannon et al. (1998) found a significant degradation in speech perception with simulated pitch shift in a four- channel AM. Dorman et al. (1997a) found that, with simulations equated to insertion depths of 22 or 23 mm, NH listeners showed reduced performance in vowel, consonant and sentence recognition. However, Rosen et al. (1999) found that the reduction in performance associated with the upward frequency transposition could be reduced by lengthy exposure to simulations. The authors found marked effect of pitch shift on AM performance in word and sentence recognition. However, the study used a four-channel implant which makes generalisation to higher number of channels used in the present study problematic.

Throckmorton and Collins (2002) described an AM of channel interaction and also other spectral anomalies that are associated with CI use, such as pitch reversals. The authors developed AMs of different aspects of electrical/neural interface signal

distortions, including pitch reversals, indiscriminable electrodes and forward masking. They compared sentence and consonant recognition abilities between the different AMs to determine which might have the greatest impact on speech perception abilities. The authors found that models of spectral channel interaction had the greatest detrimental effect on consonant recognition.

Other authors have evaluated performance with different degrees of spectral

smearing, which can be taken as a method of modelling channel interaction, at least in its spectral aspect. Shannon et al. (1998) used a simulation with overlap of filter skirts

of the noise bands, thus creating an effective spectral smearing effect. They found that channel overlap made little difference to speech recognition. However, it is worth noting that the AM they used had only four spectral channels, which means that spectral information was highly limited even without overlap.

Two identified studies to date have attempted to compare different AMs against CI user performance directly. Fu and Nogaki (2005) compared number of channels with changes in spectral resolution using spectral smearing. The outcome measure used was release from masking as shown by sentence recognition in noise. The 10 CI subjects used a variety of CI devices. AMs were based on a fixed-channel strategy using IIR filterbanks (as usual in AM studies) and varied by channel number (16, 8 and 4) and spectral overlap between channels (24 dB/octave or 6 dB/octave

slope).The authors found that release from masking in sentence recognition was modelled best by AMs in noise with broadly overlapping filters (6 dB/octave slope), although better CI users’ performance was approximated with either an 8-channel or 16-channel AM and, worse users, by a 4-channel AM. However, it should be noted that the CI users were a heterogenous group from the point of view of CI processing used and, also, that the AMs used were not based on the specific processing details of a particular device.

Laneau et al. (2006) undertook a series of experiments in which spectral overlap between adjacent channels was systematically varied. The authors were interested in perception of fundamental frequency (F0) rather than consonant recognition, but the paper is of particular interest in its use of an AM based in detail on a specific device, the Nucleus 24, implementing a specific processing strategy, ACE, and where a comparison between AM and equivalent CI user performance was made. The authors used an AM with noise band carrier stimuli. They compared pitch discrimination abilities as a function of degree of carrier overlap varying from no overlap to overlap equivalent to 10mm spread of excitation. The precise pattern of filter overlap was based on the model of channel interaction of Black and Clark (1980) and assumed asymmetric spread of excitation as noted in 2.4.1. Two separate experiments showed a close match between Nucleus 24 users and AMs with 1mm spread of excitation.A further noteworthy characteristic of this study was that the AM used the same

was compared (Laneau et al., 2004). This made the comparison between AM and CI data much more powerful than with other studies where a precise match between characteristics was not obtained, where hetereogenous groups of CI users were used, and where attempts to model electrical/neural interface factors did not have a specific physiological basis.

2.5.4. Overview of state of knowledge and knowledge gaps

• AMs have been found to be highly predictive of performance trends in channel number, although not absolute magnitude of performance levels.

• Choice of carrier stimulus probably does have some effect on AM results although it is unclear which carrier stimulus type would provide the best match/predictor of CI user performance.

• It is probable that a sine wave AM should over-predict frequency and

periodicity resolution abilities in CI users, as compared to a noise band model.

• The majority of AM studies have sought to mimic general processing

principles, rather than the fine details of processing in a specific device. Most studies have developed models based on fixed-channel processing with a IIR filter approach. This means that there is little data of direct relevance to users of the Nucleus 24 device given that this device uses an FFT filterbank and the majority of users access a peak-picking processing strategy.

• One study to date (Laneau et al., 2006) has attempted to mimic specific processing of a particular device AND aspects of the electrical/neural

interface, although the study looked at F0 discrimination in vowels rather than consonant recognition. The authors found that CI user performance was well approximated by an AM in which channel overlap was equivalent to 1 mm spread of excitation.