Speech recognition in normal hearing and sensorineural hearing loss as a function of the number of spectral channels

Benefit of high-rate envelope cues in vocoder processing: effect of number of channels and spectral region

In cochlear implants, or vocoder simulations of cochlear implants, the transmission of envelope cues at high rates (related to voice fundamental frequency, f0) may be limited by the widths of the filters used to form the channels and/or by the cutoff frequency, f(lp), of the low-pass filters used for envelope extraction. The effect of varying f(lp) in tone and noise vocoders was investigated for channel numbers, N, from 6 to 18. As N increased, the widths of the channels decreased. The value of f(lp) was 45 Hz (envelope or "E" filter), or 180 Hz (pitch or "P" filter). The following combinations of cutoff frequencies were used for channels below and above 1500 Hz, respectively: EE, PE, EP, and PP. Results from a competing-talker task showed that the tone vocoder led to better intelligibility than the noise vocoder. The PP condition led to the best intelligibility and the EE condition to the worst. For N=6, intelligibility was better for condition PE than for condition EP. For N=18, the reverse was true. The results indicate that the channel bandwidths can compromise the transmission of f0-related envelope information, and suggest that vocoder simulations of cochlear-implant processing have limitations.

Effects of moderate cochlear hearing loss on the ability to benefit from temporal fine structure information in speech

Speech reception thresholds (SRTs) were measured with a competing talker background for signals processed to contain variable amounts of temporal fine structure (TFS) information, using nine normal-hearing and nine hearing-impaired subjects. Signals (speech and background talker) were bandpass filtered into channels. Channel signals for channel numbers above a "cut-off channel" (CO) were vocoded to remove TFS information, while channel signals for channel numbers of CO and below were left unprocessed. Signals from all channels were combined. As a group, hearing-impaired subjects benefited less than normal-hearing subjects from the additional TFS information that was available as CO increased. The amount of benefit varied between hearing-impaired individuals, with some showing no improvement in SRT and one showing an improvement similar to that for normal-hearing subjects. The reduced ability to take advantage of TFS information in speech may partially explain why subjects with cochlear hearing loss get less benefit from listening in a fluctuating background than normal-hearing subjects. TFS information may be important in identifying the temporal "dips" in such a background.

Spectral and temporal cues for speech recognition: implications for auditory prostheses

Features of stimulation important for speech recognition in people with normal hearing and in people using implanted auditory prostheses include spectral information represented by place of stimulation along the tonotopic axis and temporal information represented in low-frequency envelopes of the signal. The relative contributions of these features to speech recognition and their interactions have been studied using vocoder-like simulations of cochlear implant speech processors presented to listeners with normal hearing. In these studies, spectral/place information was manipulated by varying the number of channels and the temporal-envelope information was manipulated by varying the lowpass cutoffs of the envelope extractors. Consonant and vowel recognition in quiet reached plateau at 8 and 12 channels and lowpass cutoff frequencies of 16 Hz and 4 Hz, respectively. Phoneme (especially vowel) recognition in noise required larger numbers of channels. Lexical tone recognition required larger numbers of channels and higher lowpass cutoff frequencies. There was a tradeoff between spectral/place and temporal-envelope requirements. Most current auditory prostheses seem to deliver adequate temporal-envelope information, but the number of effective channels is suboptimal, particularly for speech recognition in noise, lexical tone recognition, and music perception.

Combined Effects of Frequency Compression-Expansion and Shift on Speech Recognition

OBJECTIVE

To explore combined acute effects of frequency shift and compression-expansion on speech recognition, using noiseband vocoder processing.

DESIGN

Recognition of vowels and consonants, processed with a noiseband vocoder, was measured with five normal-hearing subjects, between the ages of 27 and 35 yr. The speech signal was filtered into 8 or 16 analysis bands and the envelopes were extracted from each band. The carrier noise bands were modulated by the envelopes and resynthesized to produce the processed speech. In the baseline matched condition, the frequency ranges of the corresponding analysis and carrier bands were the same. In the shift only condition, the frequency ranges of the carrier bands were shifted up or down relative to the analysis bands. In the compression and expansion only conditions, the analysis band range was made larger or smaller, respectively, than the carrier band range. By applying the shift to carrier bands and compression or expansion to analysis bands simultaneously, the combined effects of the two spectral distortions on speech recognition were explored.

RESULTS

When the spectral distortions of compression-expansion or shift were applied separately, the performance was reduced from the baseline matched condition. However, when the two spectral degradations were applied simultaneously, a compensatory effect was observed; the reduction in performance was smaller for some combinations compared to the reduction observed for each distortion individually.

CONCLUSIONS

The results of the present study are consistent with previous vocoder studies with normal-hearing subjects that showed a negative effect of spectral mismatch between analysis and carrier bands on speech recognition. The present results further show that matching the frequency ranges of 1 to 2 kHz, which contain important speech information, can be more beneficial for speech recognition than matching the overall frequency ranges, in certain conditions.

Simulating listener errors in using genetic algorithms for perceptual optimization

The genetic algorithm (GA) was previously suggested for fitting hearing aid or cochlear implant features by using listener's subjective judgment. In the present study, two human factors that might affect the outcome of the GA when used for perceptual optimization were explored with simulations. Listeners with varying sensitivity in discriminating sentences of different intelligibility and with varying error rates in entering their judgment to the GA were simulated. A comparison of the simulation results with the results from human subjects reported by Başkent et al. Ear Hear. 28(3) 277-289 (2007) showed that these factors could reduce the performance of the GA considerably.

Using genetic algorithms with subjective input from human subjects: implications for fitting hearing aids and cochlear implants

OBJECTIVE

To present a comprehensive analysis of the feasibility of genetic algorithms (GA) for finding the best fit of hearing aids or cochlear implants for individual users in clinical or research settings, where the algorithm is solely driven by subjective human input.

DESIGN

Due to varying pathology, the best settings of an auditory device differ for each user. It is also likely that listening preferences vary at the same time. The settings of a device customized for a particular user can only be evaluated by the user. When optimization algorithms are used for fitting purposes, this situation poses a difficulty for a systematic and quantitative evaluation of the suitability of the fitting parameters produced by the algorithm. In the present study, an artificial listening environment was generated by distorting speech using a noiseband vocoder. The settings produced by the GA for this listening problem could objectively be evaluated by measuring speech recognition and comparing the performance to the best vocoder condition where speech was least distorted. Nine normal-hearing subjects participated in the study. The parameters to be optimized were the number of vocoder channels, the shift between the input frequency range and the synthesis frequency range, and the compression-expansion of the input frequency range over the synthesis frequency range. The subjects listened to pairs of sentences processed with the vocoder, and entered a preference for the sentence with better intelligibility. The GA modified the solutions iteratively according to the subject preferences. The program converged when the user ranked the same set of parameters as the best in three consecutive steps. The results produced by the GA were analyzed for quality by measuring speech intelligibility, for test-retest reliability by running the GA three times with each subject, and for convergence properties.

RESULTS

Speech recognition scores averaged across subjects were similar for the best vocoder solution and for the solutions produced by the GA. The average number of iterations was 8 and the average convergence time was 25.5 minutes. The settings produced by different GA runs for the same subject were slightly different; however, speech recognition scores measured with these settings were similar. Individual data from subjects showed that in each run, a small number of GA solutions produced poorer speech intelligibility than for the best setting. This was probably a result of the combination of the inherent randomness of the GA, the convergence criterion used in the present study, and possible errors that the users might have made during the paired comparisons. On the other hand, the effect of these errors was probably small compared to the other two factors, as a comparison between subjective preferences and objective measures showed that for many subjects the two were in good agreement.

CONCLUSIONS

The results showed that the GA was able to produce good solutions by using listener preferences in a relatively short time. For practical applications, the program can be made more robust by running the GA twice or by not using an automatic stopping criterion, and it can be made faster by optimizing the number of the paired comparisons completed in each iteration.