Cochlear Implants

Effects of Noise and Noise Suppression on Speech Perception by Cochlear Implant Users several channels of electrical stimulation (National Institutes of Health Consensus Development Conference, Irving Hochberg, PhD; Arthur Boothroyd, PhD; Mark Weiss, MSEE; Sharon Hellman, MS Center for Research in Speech and Hearing Sciences, City University of New York Graduate Center, New York, New York

ABSTRACT The recognition of phonemes in consonant-vowel-consonant words, presented in speech-shaped random noise, was measured as a function of signal to noise ratio (S/N) in 10 normally hearing adults and 10 successful adult users of the Nucleus cochlear implant. Optimal scores (measured at a S/N of +25 dB) were 98% for the average normal subject and 42% for the average implantee. Phoneme recognition threshold was defined as the S/N at which the phoneme recognition score fell to 50% of its optimal value. This threshold was -2 dB for the average normal subject and +9 dB for the average implantee. Application of a digital noise suppression algorithm (INTEL) to the mixed speech plus noise signal had no effect on the optimal phoneme recognition score of either group or on the phoneme recognition threshold of the normal group. It did, however, improve the phoneme recognition threshold of the implant group by an average of 4 to 5 dB. These findings illustrate the noise susceptibility of Nucleus cochlear implant users and suggest that singlechannel digital noise reductiontechniques may offer some relief from this problem. (Ear Hear 13 4263-271)

COCHLEAR IMPLANTS HAVE become a viable option for the prosthetic management of profound sensorineural hearing loss. Many deaf individuals who are unable to benefit from acoustic amplification can now be given significant access to sound patterns via electrical stimulation of the auditory nerve. The benefits of cochlear implantation, however, are not uniform, and it is not possible to predict which subjects will do well with an implant and which will obtain only minimal benefit. Two significant factors have emerged from clinical experience, however. First, the most successful implantees are likely to be subjects who lost their hearing after having acquired spoken language. Second, the most successful implants are likely to be those that use Ear and Hearing, Vol. 13, No. 4,1992

1988). The most widely used multichannel implant, at the time of writing, is that developed at the University of Melbourne and manufactured by the Nucleus company. This device uses a feature extraction approach to stimulus coding with the aim of optimizing the perception of speech (Tong, Blarney, Dowell, & Clark, 1983). Relevant features of the incoming acoustic pattern are extracted and used to control various features of the pattern of electrical stimulation. The mapping of acoustic features to electrical features is done according to the response characteristics of the individual subject. In the original version of the speech processor (FO/F2), the fundamental frequency of voiced sounds controlled the frequency of electrical pulses, whereas an estimate of the frequency of the second vocal tract formant controlled the locus of stimulation. A later modification (FO/Fl/F2)added coding of the first vocal tract formant, and the most recent version (MSP) uses three channels to code the intensity of three bands of highfrequency energy (Skinner, Holden, Holden, Dowell, Seligman, Brimacombe, & Beiter, 1991). The most successful users of the Nucleus implant have attained high scores on tests involving the open set recognition of speech by implant alone. Many are even able to hold unstructured telephone conversations. (Cohen, Waltzman, & Shapiro, 1989;Dowell, Seligman, Blarney, & Clark, 1987;Skinner et al, 1991;Tye-Murray, Lowder, & Tyler, 1990). A common complaint of users of the Nucleus implant, even the most successful ones, is that performance deteriorates rapidly with increasing levels of background noise. In this respect, implantees are similar to hearing aid users in that they may be unable to perceive speech at signal to noise ratios (S/N) that are quite tolerable to normally hearing persons (Plomp, 1978). For both aided subjects and implanted subjects, alleviation of the interfering effectsof background noise would greatly expand the range of contexts in which full benefit is attained. The goal of providing effective noise reduction in hearing aids has proved elusive. Some commercial instruments do incorporate adaptive analog noise suppression circuits that reduce low-frequency ampli0196/0202/92/1304-0263$03.00/0EARAND HEARING Copyright Q 1992 by Williams 8 Wilkins Printed in the U.S.A.

Effectsof Noise

265

fication for high input levels, but empirically determined benefits have been marginal at best (Pfeffer & Ross, 1987; Van Tasell, Larsen, & Fabry, 1988). A rapidly developing digital technology has introduced new opportunities for single-channel noise suppression, but they too have failed to show the desired benefits. An example of such an approach is the INTEL system of Weiss and Aschkenasy ( 1974). This system involves the transformation of the acoustic waveform to a spectral-like domain, subtraction of the best estimate of the noise transform, and reconversion to the temporal domain. Physical measures of S/N are improved significantly by this technique, and the processed signals are noticeably less noisy and, therefore, easier to listen to than are the unprocessed signals. No improvements of intelligibility have been found, however, either for normally hearing subjects (Neuman, Mills, & Schwander, 1985) or for subjects with sensorineural hearing loss (A. Neuman, M. Bakke, T. Schwander, M. Weiss, & H. Levitt, submitted). These negative findings can be attributed in part to the fact that the algorithm tends to remove low-energy speech sounds (which tend to be the higher frequency sounds) along with the noise. There is, however, an important difference between hearing aids and feature extraction systems such as that used in the Nucleus implant. Specifically, the hearing aid transmits an acoustic signal, consisting of both speech and noise energy, to the damaged ear of the user. Although this signal may be changed in some way by amplification and processing, it still represents an acoustic code, and the extraction of information from this code relies, in part, on the integrity of the cochlea. In the Nucleus implant, however, the extraction of information is the responsibility of electronic circuits within the processor. The resulting patterns may be distorted because of the presence of noise in the input, but the noise does not appear as additional random energy in the electrical output. This difference raises the possibility that a single-channel noise reduction system such as INTEL might improve the performance of cochlear implants by reducing the error rate of the feature extraction circuits, enabling them to pass on fewer errors in the patterns of electrical stimulation. Empirical support for the notion that the INTEL system might reduce noise susceptibility in Nucleus implantees comes from work on the automatic recognition of speech. Cupples (1984) has found improvements of 30 to 50% on measures of word recognition when noisy speech is preprocessed using the INTEL algorithm. There is, however, no guarantee that digital noise suppression systems such as INTEL will improve feature extraction in a system such as the Nucleus processor. The interaction between the noise reduction process, which is inevitably nonlinear, and the feature extraction circuitry, which is also nonlinear, is complex, and outcomes are difficult to predict. Also, even if physical improvements of performance can be demonstrated, it is not clear how large they would need to be in order to provide significant improvements of 264

Hochberg et al

speech perception for the implant user. The general goal of the study to be reported was to test the hypothesis that single-channel digital noise suppression would reduce the interfering effects of noise for users of a feature extraction cochlear implant. The specific goals were: (1) To determine, relative to normal hearing individuals, the effect of noise on the speech perception performance of users of the Nucleus cochlear implant, and (2) to determine the extent to which this effect might be alleviated, using the INTEL noise reduction algorithm at the input to the implant’s own processor. I

METHOD

Subjects Cochlear Implant Users Ten postlingually deafened adults who were highly successful users of the Nucleus cochlear implant served as subjects. Data on the implantees are given in Table 1. The group consisted of five males and five females who ranged in age from 25 to 72 yr, with a mean age of 45.5 yr. The duration of profound deafness ranged from less than 1 yr to 45 yr, with a mean of 13 yr. The age at which subjects received their Nucleus implant ranged from 23 to 70 yr, with a mean age of 43 yr, and the duration of experience with this device ranged from 6 to 5 1 mo, with a mean of 24 mo. A criterion of subject selection was a high level of success with the implant. Operationally, this was defined as a phoneme recognition score, by implant alone, in excess of 20% when measured with the NU-6 word lists. In fact, scores for these subjects ranged from 24 to 56% with a mean of 33.9%. Eight of the subjects used the Nucleus MSP processor, which, it will be recalled, encodes estimatesof fundamentalfrequency (FO) and the frequencies of the first and second vocal tract formants (F1 and F2) together with the energy content of three high-frequency spectral regions (Skinner et al, 1991).

The other two subjects used the WSP processor, which encodes only FO, F1, and F2. Normal-Hearing Subjects Ten normal-hearing subjects also participated in the study. The group consisted of four males and six females who ranged in age from 26 to 42 yr, with a mean age of 34.2 yr, and had normal audiometric hearing levels. Test Stimuli Speech Speech perception performance was measured as the percent recognition of phonemes in lists of consonant-

vowel-consonant words. The word lists used for this study were the AB isophonemicword lists (Boothroyd, 1984). These materials consist of 15 different lists, each containing 10 words. The 10 words are constructed from 30 phonemes that remain constant from list to list, with each phoneme occumng only once in each list. The materialswere designed to facilitate the generation of performance/intensity functions in a reasonably short period of time. The recording of the AB lists used in this study was of a female talker and was already available on a video laser-disc (Boothroyd, Hnath-Chisolm, Hanin, & Kishon-Rabin, 1988). Presentation of the words and scoring of responses were under computer control. Noise For the purposes of this study, the speech signal was mixed with speech spectrum-shapednoise (constantspectrum level below 800 Hz and 10 dB/octave role-off above 800 Hz). S/N was defined as the difference,in decibels, between Ear and Hearing, Vol. 13, No. 4,1992

Table 1. Information on cochlear implantees.

Subject Number 1

Age and sex Age when implanted (yr) Years of profound deafness Etiology

2

3

4

6

5

7

8

9

35 F 31

42 F 39

62 F 61

65 F 62

32 M 30

41 M 38

72 M 70

46

25 M 23

31 F 31

10

10

35

Effects of noise and noise suppression on speech perception by cochlear implant users.

The recognition of phonemes in consonant-vowel-consonant words, presented in speech-shaped random noise, was measured as a function of signal to noise...
992KB Sizes 0 Downloads 0 Views