Improved neural representation of vowels in electric stimulation using desynchronizing pulse trains

Abstract

Current cochlear implant processors poorly represent sound waveforms in the temporal discharge patterns of auditory-nerve fibers (ANFs). A previous study [Litvak et al., J. Acoust. Soc. Am. 114, 2079-2098 (2003)] showed that the temporal representation of sinusoidal stimuli can be improved in a majority of ANFs by encoding the stimuli as small modulations of a sustained, high-rate (5 kpps), desynchronizing pulse train (DPT). Here, these findings are extended to more complex stimuli by recording ANF responses to pulse trains modulated by bandpass filtered vowels. Responses to vowel modulators depended strongly on the discharge pattern evoked by the unmodulated DPT. ANFs that gave sustained responses to the DPT had period histograms that resembled the modulator waveform for low (< 5%) modulation depths. Spectra of period histograms contained peaks near the formant frequencies. In contrast, ANFs that gave a transient (< 1 min) response to the DPT poorly represented the formant frequencies. A model incorporating a linear modulation filter, a noisy threshold, and neural refractoriness predicts the shapes of period histograms for both types of fibers. These results suggest that a DPT-enhanced strategy may achieve good representation of the stimulus fine structure in the temporal discharge patterns of ANFs for frequencies up to 1000 Hz. It remains to be seen whether these temporal discharge patterns can be utilized by cochlear implant subjects.

FIG. 1

(a) Implementation of the DPT-enhanced stimulation strategy. The left panel shows two pulse train stimuli that might be generated by a cochlear implant processor: one, strongly modulated, is the CIS signal, and the other, unmodulated, is the DPT. We assume that the DPT-enhanced stimulus can be modeled by a carrier of the same frequency as the DPT that is more weakly modulated than the original CIS stimulus (right). (b) Schematic representation of the electric stimuli used in this study. The stimulus consists of a sustained DPT, which is modulated every second for 400 ms. The modulation waveform and/or modulation depth are changed on each successive modulated segment. The entire cycle of modulation waveforms and modulation depths repeats for up to 10 min.

FIG. 2

Block diagram of the stochastic threshold model for predicting ANF responses to arbitrary modulation waveforms applied to the DPT. The model takes as input the modulation waveform m(t), which is passed through a modulation filter. It produces a spike whenever the filtered input crosses a noisy threshold. The threshold is the sum of a Gaussian noise term n(t) and a deterministic term θ(t) which depends on the time since the previous spike. The model outputs the set of spike times {t_i}. Free parameters of the model are the modulation filter, the resting threshold θ₀, and the noise standard deviation σ.

FIG. 3

Period histograms of responses to vowel modulators for a sustained DPT responder (pseudo-spontaneous rate: 28 spikes/s). Top row: waveforms of the four filtered vowel modulators. Lower four rows: Period histograms of responses to each vowel for four modulation depths. The ordinate in each panel is in spikes/s. The numbers in each panel indicate the average discharge rate during the vowel modulator in spikes/s.

FIG. 4

Period histograms of responses to the vowel modulators for a transient DPT responder (pseudo-spontaneous rate: 0 spike/s). Same format as in Fig. 3.

FIG. 5

Maximum correlation coefficient between the period histogram of ANF responses and the vowel waveform as a function of pseudo-spontaneous rate for 2.5% (top) and 10% (bottom) modulation depths. Symbols code the correlation coefficients for each vowel. Data are only included if the response contains at least ten spikes and the bootstrap 90% confidence intervals for the correlation coefficient are below 0.1.

FIG. 6

Average of period histograms for two vowel modulators. The top row shows the spectra of the two vowels. Peaks in these spectra occur at the harmonics of the 100-Hz fundamental that are closest to the a formant frequency: 200 Hz for low-pass /u/, and 800 Hz for high-pass /u/. The bottom two rows show the average spectra of period histograms in response to both vowels for modulation depths of 2.5% (left) and 10% (right). The spectra were averaged separately for transient DPT responders (pseudo-spontaneous rates below 5 spikes/s; middle row), and sustained responders (bottom row). Numbers in each panel indicate the number of fibers whose responses are included in the average. Responses had to contain at least ten spikes to be included. The ordinates of each panel are in spikes/s.

FIG. 7

Pooled autocorrelation histograms for two vowel modulators. The top row shows the autocorrelation functions of both vowels. Dashed lines show multiples of the stimulus period. The bottom two rows show pooled autocorrelation histograms of the electric responses for modulation depths of 2.5% (middle row) and 10% (bottom row). Numbers in each panel indicate the number of fibers whose responses were included in the pooled autocorrelation histogram. Responses had to contain at least ten spikes to be included. Ordinates in each panel are numbers of intervals per bin.

FIG. 8

Neural MTFs estimated from responses to vowel modulators at 1% modulation depth for three fibers with pseudo-spontaneous rates below 130 spikes/s (left), and three with pseudo-spontaneous rates above 130 spikes/s (right). Each panel shows the neural MTF magnitudes estimated from responses of one fiber whose pseudo-spontaneous rate is indicated inside the panel. Symbols code estimates based on different vowels. The MTF could only be estimated at frequencies for which the vowel stimulus has energy. Error bars represent the 99% confidence intervals computed by bootstrapping ANF spike trains (see Sec. II). For the three fibers on the right, the intrinsic periodicities (defined as the inverse of the largest mode in the interval histogram of responses to the unmodulated DPT) were, from top to bottom, 234, 537, and 383 Hz.

FIG. 9

Magnitude (top) and phase (bottom) of the average neural MTF estimated from responses to vowel modulators with 1% modulation depth. The black solid line shows the grand average MTF, which was computed by averaging individual MTFs across the four vowel modulators for ANFs with pseudo-spontaneous rates above 130 spikes/s. The symbols show the average MTF based on responses to individual vowels. The stars indicate the average, and the error bars 99% confidence intervals of the MTF magnitude estimated from responses to sinusoidal modulators with 1% modulation depth (Litvak et al., 2003a).

FIG. 10

Period histograms of responses of a model fiber with a high pseudo-spontaneous rate (50 spikes/s, θ₀ / σ = 2.2) to the vowel stimuli. Same format as in Fig. 3. The second number in each panel represents is the stimulus-response correlation as defined as in Fig. 5, while the first number is the average discharge rate during the vowel stimulus.

FIG. 11

Period histograms of responses of a model fiber with a low pseudo-spontaneous rate (0.0014 spikes/s, θ₀ / σ = 5) to the vowel stimuli. Same format as in Fig. 10.

FIG. 12

Correlation coefficient between the period histograms of STM responses to vowel modulators and those of ANF responses as a function of pseudo-spontaneous discharge rate for modulation depths of 2.5% (top) and 10% (bottom). Each data point shows the correlation coefficient between the period histogram for one ANF and the histogram for a model fiber having the same pseudo-spontaneous rate. To account for differences in neural delay across fibers, we allowed a ±0.4-ms shift between the two responses. Symbols code responses to different vowels. Data are only included if the response contains at least ten spikes and the bootstrap 90% confidence intervals for the correlation coefficient are below 0.1.