- Email: [email protected]
Preliminary results of a novel enhancement method for high-frequency hearing loss
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 2 ( 2 0 1 1 ) 277–287
journal homepage: www.intl.elsevierhealth.com/journals/cmpb
Preliminary results of a novel enhancement method for high-frequency hearing loss Umut Arioz a,∗ , Kemal Arda b , Umit Tuncel c a b c
Medical Informatics, Informatics Institute, M.E.T.U, Turkey Radiology, Ankara Oncology Education and Research Hospital, Turkey Otorhinolaryngology, Ankara Oncology Education and Research Hospital, Turkey
a r t i c l e
i n f o
a b s t r a c t
Article history:
In this study, a software program was developed for high-frequency hearing loss subjects
Received 11 September 2009
that includes a detailed audiogram and novel enhancement methods. The software per-
Received in revised form
forms enhancements of the audibility of high-frequency sounds according to the subject’s
13 May 2010
detailed 31-point audiogram. This provides subject-specific gains in the entire frequency
Accepted 13 May 2010
spectrum, and especially for high frequencies, of sounds. Amplification, compression, and transposition are the three main processing methods used to obtain the desired enhance-
Keywords:
ments for the subjects. For low frequencies, only the amplification method was used
Amplification
according to the dB value of the input. For mid and high frequencies, the compression and
Compression
transposition methods were used together. To obtain the preliminary results of the study,
Transposition
10 subjects were enrolled in a detailed audiogram study for five weeks. In the study, envi-
Audiogram
ronmental, music, and speech sounds were used. While the perceptual mean performances
High frequency
of the subjects were in the range of 25.33–63.77% in the first week, those values increased
Hearing loss
to 68.75–95.75% in the fifth week. In particular, all noisy and speech sounds were more significantly identified and understood by the subjects with the enhancement method (from 25.33% to 87.5% and from 42.33% to 90.5%, respectively). Three subjects had dropped out at the end of the study and small number of participants are the limitations of this study; however, as a preliminary result, some ideas can be inferred from the results for a smaller set of subjects obtained in the five weeks of the study. Thus, the reliability of the study needs to be tested with more subjects and a comparison with their own hearing aids should be conducted. © 2010 Elsevier Ireland Ltd. All rights reserved.
1.
Introduction
Many hearing-impaired individuals have a greater loss of hearing sensitivity at high frequencies than at low frequencies. It has been shown that high-frequency acoustic cues impact speech-recognition abilities even for normal-hearing children [1]. Children with hearing loss are more affected by limited high frequencies in terms of speech recognition [2], and may
∗
also exhibit errors in morphology such as plural markers and verb tense [3]. Providing high-frequency audibility to subjects with severe high-frequency hearing loss presents several challenges for the clinician when fitting a hearing aid. Firstly, the level difference between discomfort and the threshold of audibility is usually much less than that of a normally hearing ear [4]. Secondly, acoustic feedback will often limit the amount of usable gain, depending on individual character-
Corresponding author at: Medical Education and Informatics, Hacettepe University, Sihhiye 06100, Ankara, Turkey. Tel.: +90 312 3052578. E-mail address: [email protected] (U. Arioz). 0169-2607/$ – see front matter © 2010 Elsevier Ireland Ltd. All rights reserved. doi:10.1016/j.cmpb.2010.05.004
278
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 2 ( 2 0 1 1 ) 277–287
istics such as how effectively the ear canal is sealed by the earmold. Up to now, researchers have reported the results of trying many different enhancement methods for high-frequency hearing losses in the literature. Amplification, compression, and transposition are the most frequently used techniques. However, all three methods have both advantages and disadvantages related to different parts of the frequency spectrum. Although conventional amplification can provide usable low-frequency information, the amplified high-frequency sounds of speech are often inaudible due to the severity and configuration of the hearing loss. Until recently, the findings have suggested that an increase in high-frequency gain may not improve, and in some cases may even degrade, speech recognition for listeners with high-frequency hearing losses [5,6]. Other investigators have shown some improvements in specific subjects (e.g., flat hearing loss) or under certain conditions (e.g., soft speech and fricatives) [7,8]. It is recommended that clinicians use caution in providing amplification in the high frequencies when the hearing loss in this region is greater than approximately 55 dB [9,10]. There is a need, therefore, to consider alternative sound-processing options for individuals with severe high-frequency losses. Ideally, this processing will preserve useful low-frequency amplification and provide some additional high-frequency information. An alternative approach might be the use of frequency compression or transposition schemes, whereby high-frequency signals are shifted to lower frequencies to provide adequate audibility. This type of approach has produced mixed results, with some studies showing substantial improvement and others showing no improvement or degradation in performance [11–13]. With frequency compression, high-frequency sounds are compressed into lower frequencies where there is better residual hearing and less amplification is required. This allows for a decrease in gain in the high frequencies where the subject’s hearing is more damaged. The decrease in gain will also reduce feedback, distortion, and discomfort. The lack of spectral overlap between the shifted and the unshifted signals can be identified as an advantage of frequency compression. On the other hand, that it does not preserve frequency ratios for high frequencies can be identified as a disadvantage of frequency compression. The perceptual performance of a number of linear and nonlinear frequency-compression schemes was evaluated in one study [14]. In a preliminary study, six subjects with normal hearing participated in experiments that investigated whether any of the schemes could improve the discriminability of consonant stimuli. Although none of the schemes provided better performance than a standard condition that applied only low-pass filtering to the stimuli, the best scheme was found to be a variant that progressively increased the amount of frequency compression for input frequencies above approximately 1200 Hz. In another study [15], the quality and intelligibility of speech sounds were improved for hearing-impaired subjects through compression along the frequency axis using an FFT-based approach. In this technique, the signal for each frame was transformed from the time domain to the fre-
quency domain using FFT. After the amplitude and phase spectra of the FFT were calculated, a have attained speech perception benefits when listening to proportional frequencycompression spectrum was computed for each band. Then, the processed spectrum was multiplied by the original phase spectrum to re-synthesize a band-limited signal. Finally, with the overlap add technique and IFFT process, the final signal was obtained by some listeners [13]. An advantage of this method is that the frequency ratios are preserved. These ratios may be particularly important cues for the recognition of vowels in speech [16]. In a recent study [17], nonlinear frequency compression was used with 1.6 kHz cutoff frequency and 0.5 compression exponent parameters. This research provided better recognition of monosyllabic words and achieved greater high-frequency sensation levels by enabling frequency compression than with conventionally fitted hearing aids. Various sound-processing schemes have been developed over the past decades that have attempted to present information from high-frequency regions of speech at lower frequencies. In extreme cases, frequency lowering (“frequency shifting” or “transposition”) may be the only way to provide this extra information acoustically. One scheme [18] implementing disproportionate frequency shifting, and several studies using early types of frequencylowering methods [19,20] reported little success in providing speech understanding benefits. This may have occurred for a variety of reasons. Much of the information related to the spectral shape of the incoming signal was lost as a result of the processing technique. These schemes may have provided additional high-frequency information at the expense of other perceptual cues by overlapping the shifted and unshifted signals. Some early attempts at frequency transposition converted signals in the 3–6 kHz range into low-frequency noise below 1.5 kHz by passing signals through a nonlinear modulator [18,21]. No significant improvements in speech intelligibility were found with the device [19,20]. Proportional frequency shifting, using a ‘slow play’ method, is an alternative sound-processing technique [22,23]. Segments of the speech signal are recorded and then played back at a slower speed than employed for recording. The TranSonic (AVR Sonovations Inc., USA, 1991) and ImpaCt DSR675 (AVR Sonovations Inc., USA, 1998) are hearing instruments that incorporate such processing. Incoming signals dominated by components at frequencies above 2.5 kHz are shifted down by a factor that is programmable for each listener. Positive outcomes were reported when the TranSonic was fitted to a small number of hearing-impaired children [24]. A similar study with adults demonstrated speech perception benefits with the device [23,25], and reported that the amplification characteristics of the TranSonic in the low frequencies, rather than its frequency shifting characteristics, may have provided most of the benefit. In another transposition study for children [11], the system categorized the incoming signal according to whether its energy peak was above or below 2.5 kHz. Signals below 2.5 kHz were categorized as vowels, and their transposition coefficients were in the range of 1–1.4; signals above 2.5 kHz were categorized as consonants, and their transposition coef-
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 2 ( 2 0 1 1 ) 277–287
ficients were in the range of 1–6. At the end of the study, only one-third of the initial group fitted was still using the device and obtaining improvements in the specific frequencies. In a more successful study [26], an audibility extender (AE) was employed as a frequency-lowering technique that uses linear frequency transposition to move inaudible highfrequency sounds to audible low-frequency regions. First, the algorithm receives information about the wearer’s hearing loss and decides which frequency region (source octave) will be transposed by the help of the AE’s signal processing component, which is called a dynamic integrator. The frequency at which transposition begins is called the start frequency. Typically, in which system, one octave of sounds above the start frequency was transposed. The linearly transposed signals were mixed with the original signal below the start frequency as the final output. As a result, subjects with a sloping highfrequency hearing loss subjectively preferred the AE for its much simpler stimulus. In a more recent study [27], an edge frequency (fe) parameter was selected as an edge frequency for the dead region. A band from 1 fe to 1.7 fe was selected as the destination for the transposed frequency components. High-frequency components from a source band well within the dead region (2–2.7 fe) were shifted into this band, but no frequency compression was applied. In the algorithm used, frequencies below fe were left unchanged. Transposition only occurred if the short-term spectrum was dominated by high frequencies. The results of the tests showed that although there was no significant overall benefit to transposition, the processing did not impair consonant identification. In our study, we used all three methods (amplification, compression, transposition) in an attempt to provide highfrequency audibility in high-frequency hearing loss patients. Because each of these methods has an advantage for a specific range of the spectrum, we planned to benefit from all of the techniques. In order to do so, software was developed to perform the signal processing automatically according to the patient’s audiogram. In this way, we wanted to overcome of the problem of generalization of the same algorithm for every patient which may be the main reason for unsuccessful results of the past studies. Thus, we developed a patient-specific algorithm based on detailed audiograms.
2.
Materials and methods
2.1.
Subjects
Ten hearing-impaired adults, two women and eight men, participated in the study. Detailed audiogram measurements had been taken from all subjects at the time of testing, but three subjects did not follow the study after one week. The majority of the subjects participated in the whole study for software sound testing (seven subjects for two weeks, five subjects for three weeks, and four subjects for five weeks). The number of subjects decreased because of personal reasons. The reasons of the subjects’ nonattendance were distance between their homes and test center and not having enough mobility to come the tests (subject 5); having huge hearing loss and not expecting a benefit from the study (subjects 1 and 10);
279
and having other health problems more important than hearing loss after the study began (subjects 2, 7, and 4). Relevant information about all the subjects is provided in Table 1. The majority of the subjects had moderate-to-severe hearing loss. Their hearing threshold levels, measured conventionally with headphones by the standard method, are listed in Table 2. Nine of the subjects were experienced hearing aid users. They had worn hearing aids for several years and wore their current hearing aids on a daily basis. The remaining subject had not previously worn a hearing aid (subject 5). The subjects were not paid for their participation in the experiment, although expenses such as travel costs were reimbursed. All explanations about the study were given to the subjects at the beginning of the study. The participation of the subjects in this experiment was in accordance with the ‘Guiding principles for research involving human or animal subjects’.
2.2.
Stimuli
To evaluate the efficacy of the enhancement algorithm, the aided hearing sensitivity, speech recognition, and identification of environmental sounds and music were measured to determine whether new high-frequency cues were available. Thus, seven sounds with different dB values from 20 to 80 dB range were prepared in the categories of environmental, music, and speech sounds with their noisy forms (a total of 19 for noiseless sounds [7 environmental sounds, 7 music sounds, and 5 speech sounds] and 17 noisy sounds [6 environmental sounds, 6 music sounds, and 5 speech sounds]). All sounds were selected according to their high-frequency content ([0–8 kHz] frequency spectrum range). Environmental sounds were selected by considering the high-frequency sounds in daily life that occur most frequently, like the sound of water in the kitchen, paper tearing, jingling keys, or two plates hitting one another. Music sounds were short selections from instrumental songs. Speech sounds were selected from the Turkish film cues, and incorporated different speaking properties. Noisy sounds were constructed by adding brown noise, pink noise, or white noise to the original sounds. This was done using the signal processing’s simple addition commands in the time domain. All noises were tried one by one to obtain the most suitable noisy sound and selected with the criterion of being not too dominant over the original sound. All noise sounds were taken from the standard sounds library of the MATLAB software [28]. For all subjects, determining and understanding the type of sound was conducted in five different ways (1. without saying the sound’s property or type [undirected] for all types of sounds; identifying the sounds as: 2. noisy sounds for all types of sounds, 3. environmental sounds for both noise-free and noisy sounds, 4. music sounds for both noise-free and noisy sounds, 5. speech sounds for both noise-free and noisy sounds).
2.3.
Software
All signal processing was done using our own software. The software was developed using MATLAB [28]. Mainly, the pro-
280
Left Binaural Right Right – Left Binaural Binaural Binaural Right Linear Linear Linear Linear – Linear Linear Linear Linear Linear Fully automatic digital Fully automatic digital Fully automatic digital Fully automatic digital – Fully automatic digital Fully automatic digital Fully automatic digital Fully automatic digital Digitally programmable Bernafon F 401 Bernafon Neo 401 Bernafon Win 112 Bernafon Win 401 – Bernafon Win 102 Audio service Nova 35 Bernafon Win 112 Audibel DeluxA Bernafon AF 310 Unknown Presbyacusis Industrial noise exposure High-level noise exposure Presbyacusis Presbyacusis Unknown Unknown Unknown Unknown 51 67 59 37 53 57 28 17 22 48 Subject 1 Subject 2 Subject 3 Subject 4 Subject 5 Subject 6 Subject 7 Subject 8 Subject 9 Subject 10
M M M F M M M M F M
Probable etiology of hearing loss Sex Age Subject
Table 1 – Subjects’ information about their hearing loss.
Type of own hearing aids
Features of own hearing aids
Processing strategy of conventional hearing device
Ears fitted
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 2 ( 2 0 1 1 ) 277–287
gram has two parts. In the first part, the detailed audiograms of the patients can be taken at 31 different frequency points. In the second part, the main signal processing can be done according to each patient’s detailed audiogram. Both parts have user friendly graphical interfaces. The user can construct any patient-specific test sounds using standard embedded sounds, his/her own specified sounds, or recorded sounds. The software works according to octave calculation of the sounds. By processing with octaves, the harmonics of the sound signal is preserved and distortions after processing are minimized.
2.4.
Signal processing
The frequency spectrum was split into three parts (0–1 kHz as low frequency, 1–4 kHz as mid frequency, and 4–8 kHz as high frequency) by taking into consideration the studies mentioned in Section 1. Thus, the goal was to avoid spectral or vocal distortions in the sounds throughout the whole process. Splitting into octaves is a critical process for applying the desired part of the spectrum with suitably without distorting its characteristic properties. Otherwise, unnatural sounds and overlaps and gaps occur; these are some of the main reasons for the unsuccessful results of earlier studies. The low-frequency part consisted of 3 octaves (125–250, 250–500, 500–1000); the midfrequency part consisted of 2 octaves (1000–2000, 2000–4000); and the high-frequency part consisted of 1 octave (4000–8000). Each octave’s value was determined as an average from its audiogram dB values. Audiogram frequency points were determined as giving enough information for low frequencies and detailed information for mid and high frequencies. Low-, mid, and high-frequency regions had the following 3 ranges of points that determined the enhancement parameters: 1. 2. 3. 4. 5. 6.
octave = 125–250 – 1 measurement – 1 range octave = 250–500 – 2 measurements – 1 range octave = 500–1000 – 4 measurements – 1 range octave = 1000–2000 – 4 measurements – 1 range octave = 2000–4000 – 8 measurements – 2 ranges octave = 4000–8000 – 12 measurements – 3 ranges
For the low-frequency part, only the amplification process was applied to the sounds. This application was conducted according to the octaves’ dB value. The gain was calculated by determining the decibel difference of the input signal and patient’s hearing loss for each octave. Thus, unneeded high amplification was not applied to the subjects (for example, for a 40 dB input sound, 10 dB amplification was applied for 50 dB hearing loss in that frequency range). For the mid- and high-frequency parts, frequency compression and transposition were applied as an enhancement. Frequency compression was applied for mid frequencies, while frequency transposition was applied for high frequencies. However, the algorithm worked by taking into account both the mid- and the high-frequency parts. The frequency compression and transposition parameters (compression region, compression rate, transposition region, transposition rate) were determined automatically by the software algorithm according to the audiogram of the patient. The criteria of the determination of the enhancement parameters were
281
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 2 ( 2 0 1 1 ) 277–287
Table 2 – Hearing thresholds of the subjects. Subject
Hearing loss (dB)/frequency (kHz) Ear
0.25
0.5
1
2
4
6
8
S1
L R
15 20
30 20
60 55
70 80
80 80
85 90
90 85
S2
L R
15 15
25 20
30 30
50 55
55 60
65 70
55 70
S3
L R
43 25
35 20
65 55
70 65
90 90
90 120
90 120
S4
L R
15 20
20 20
20 20
65 60
80 75
80 85
85 80
S5
L R
10 10
10 10
10 10
20 25
65 60
65 65
S6
L R
45 20
45 30
50 30
60 30
80 40
110 60
S7
L R
45 30
60 45
65 70
65 60
65 65
60 65
S8
L R
40 40
50 60
75 70
90 100
110 90
115 90
S9
L R
15 15
30 35
45 40
45 45
55 50
55 55
S10
L R
35 35
45 35
75 70
85 80
120 120
120 120
that no gap could occur and that there was no overlap between the compressed and the transposed parts. Thus, in a case where two processes (compression and transposition) occurred, the software searched for the compression ratio and transposition amount that were suitable to avoid overlapping or gap positions after the whole process had occurred. It did so by taking into account the shape of the audiogram. The aided thresholds for our software are shown in Fig. 1 for every point of four different patients’ audiograms. Enhancements for the mid- and high-frequency parts were applied according to the dB value of the input signal. After measuring the dB value of the input signal, no calculation was needed for any enhancement of the part of the frequency spectrum that was below the dB value, as this part was also audible for the patient. Enhancements were applied only to the part of the spectrum above dB value. Thus, fewer calculations and measurements were obtained for the same enhancement using a less time-consuming process. A flow chart of the enhancement processing can be seen in Fig. 2.
2.5.
Audiogram measurement
Detailed audiograms of all patients were taken by our software through the use of headphones. For the audiogram test, standard audiological test sounds (125, 250, 500, 1000, 2000, 4000, 6000, 8000 Hz) and detailed frequency sounds (750, 1250, 1500, 1750, 2250, 2500, 2750, 3000, 3250, 3500, 3750, 4250, 4500, 4750, 5000, 5250, 5500, 6500, 7000, 7500 Hz) were used that were reconstructed from standard sounds by sampling processing. All tests were done in a test room that was sound-isolated and specially designed for audiogram tests.
2.6.
Test procedure
Subjects were specially selected for our study. Because of our test procedures, there were some constraints involved that limited the number of subjects. At the beginning of the study, the requirements and obligations of the study were explained to the subjects; if they accepted them, the study was initiated. These requirements involved being able to come to test weekly, having time to listen to the sounds at least three times a day, and being able to listen the sound CD in their home with a computer or other electronic equipment with only headphones. The latter was made compulsory to obtain the same environment for all subjects. For the tests, a detailed audiogram was taken at the first meeting. Then, according to the subject’s audiogram, all sounds were enhanced with the software and stored on a CD. At the second meeting, a few days later, the prepared patient-specific sound CD and patient listening follow-up form were given to subjects; on the same day, the first test with enhanced sounds was conducted as a control study. After this, every 7 days, testing was done with the same enhanced patient-specific sounds and follow-up form was controlled.
3.
Results
3.1.
Detailed audiogram
The audiograms of the 10 subjects taken by both the standard method and our detailed 31-point method are shown in Fig. 3. The correlation coefficients between the audiograms were in the range of 0.9482–0.9889, with a mean of 0.9682. These were calculated after adding linearly the
282
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 2 ( 2 0 1 1 ) 277–287
Fig. 1 – Aided thresholds, example of four subjects (filled triangle: after enhancement, empty triangle: before enhancement).
missing values between the standard method values. The correlation coefficients of our audiograms and the standard measurement showed not only the reliability of the usage but also the need for more frequency points in the audiogram measurement. The difference and the additional frequency information obtained can be seen in Fig. 3. In particular, the difference and the extra frequency information can be seen in the 0–2000 Hz range for subjects 1, 2, 4, and 9; in the 2000–4000 Hz range for subjects 5 and 6; and in the 4000–8000 Hz range for subjects 4 and 6. For subjects 3 and 7, the 31-point audiogram did not provide extra information because of their flat and linear hearing loss configuration. Moreover, for subjects 8 and 10, the 31-point audiogram provided extra information only in the low frequencies because of the highest level of flat hearing loss in the high frequencies.
3.2.
Aided thresholds
All subjects obtained aided thresholds; as examples, four subjects’ aided thresholds, along with their pre-processing thresholds, are shown in Fig. 1. Subjects 1 and 2 obtained aided thresholds particularly in the 1500–3000 Hz range through the automatic addition of a 20–35 dB gain. Subjects 5 and 6 got aided thresholds in particularly the 2000–5000 Hz range through the automatic addition of 25–35 dB gain.
3.3.
Subject tests
In our study, all subjects listened to the sounds and gave responses like hearing/not hearing and understanding/not understanding. The percentages of both hearing and understanding responses were calculated for each subject every
Fig. 2 – Flowchart of the algorithm.
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 2 ( 2 0 1 1 ) 277–287
283
Fig. 3 – Detailed and standard audiograms of all ten subjects (filled circle: detailed audiogram values, empty circle: standard audiogram values).
284
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 2 ( 2 0 1 1 ) 277–287
Fig. 4 – Five weeks correct understanding percentage values of undirected sounds.
week. These values are shown in Figs. 4–8. Of all the subjects, seven came to second meeting; for the other three subjects; no comparisons were obtained for this study. Only 4 subjects completed all five weeks of the study. For all subjects participating in the study, the mean, minimum, and maximum values of the results are shown in Table 3. As shown, there were small decrements for noise-free sounds in the fourth week and for speech sounds in the fifth week; for the rest of the measures, an increase can be seen very clearly and significantly from the first week to the fifth week.
4.
Discussion
The main goal of our study was to offer patient-specific fittings for hearing losses. Differences between patients include
lifestyle, speech understanding, and cognitive ability. Each of these differences may result in one patient requiring a different fitting than another patient who has similar levels of hearing loss. In addition, any difference in the audiogram must be taken into account for better satisfaction and life quality. As seen in the literature, the amplification technique is much more efficient for low frequencies. Frequency compression can be considered in the mid frequencies, where amplification does not work or gives unsatisfactory results. Finally, frequency transposition can be used for frequency lowering of high frequencies to much more audible levels. By taking exploiting the advantages of the three main methods, a new method has been developed. In our method, amplification, compression, and transposition are responsible for the 0–1000 Hz, 1000–4000 Hz, and 4000–8000 Hz regions, respectively. These regions were determined by taking into
Fig. 5 – Five weeks correct understanding percentage values of noisy sounds.
285
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 2 ( 2 0 1 1 ) 277–287
Fig. 6 – Five weeks correct understanding percentage values of environment sounds.
Fig. 7 – Five weeks correct understanding percentage values of music sounds.
Table 3 – Mean, minimum and maximum values of perceptual performances of the subjects. First week mean (min–max) Undirected Directed Noiseless Noisy Env. sounds Music Speech
30.66 (13–49) 55.55 (27–81) 63.77 (33–89) 25.33 (0–45) 35 (0–56) 56.44 (40–75) 42.33 (6–78)
Second week mean (min–max) 42.28 (25–70) 75.42 (58–90) 74.14 (57–89) 50.28 (17–80) 49.85 (30–65) 70.85 (38–90) 66 (48–90)
Third week mean (min–max) 61.8 (50–73) 88.2 (68–100) 81.8 (67–92) 67.8 (46–88) 58.6 (30–75) 83.6 (50–100) 82.2 (75–89)
consideration of the sound octave parts and characteristics of the frequency regions. Thus, the distortion of the signals was ameliorated and the speech discrimination rates were improved.
Fourth week mean (min–max) 72.5 (67–78) 94 (83–100) 78 (63–92) 82 (75–93) 60.25 (38–80) 87.75 (85–90) 92.5 (85–100)
Fifth week mean (min–max) 80.25 (77–83) 95.75 (93–100) 82 (63–92) 87.5 (82–92) 68.75 (50–80) 95 (85–100) 90.5 (83–100)
Based on these observations, the audiogram configuration is suggested to be an important factor in recommending an enhancement algorithm. The frequency at which the hearing loss becomes severe, as well as the
286
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 2 ( 2 0 1 1 ) 277–287
Fig. 8 – Five weeks correct understanding percentage values of speech sounds.
steepness of the slope of the audiogram, may be important. The most important and vital role in determining the hearing aid algorithm or fitting is the patient’s audiogram, because most of the processes in this study were carried out according to it. Thus, true measurement is very crucial. Because of the non-objective nature of audiogram measurement, future research should seek to improve the accuracy of the measurement. In this study, we tried lessen the effect of non-objectivity by obtaining the measurement for more points relating to high-frequency loss in order to produce a more accurate audiogram measurement. In the measurement, standard audiogram methods were used. Using more points than are employed for standard measurement, not only was extra information gained that standard measurement cannot detect, but also there was no missing frequency information, which is especially important for high frequencies. Using more points in the audiogram measurement provided more accurate fitting algorithms for the patients; this results in more satisfaction in terms of patients’ hearing aids. This improvement, of course, is important and specific to high-frequency loss patients and our study’s methodology. In our approach, applying a more accurate fitting algorithm to patients is more important than the time it takes to obtain a 31-point audiogram rather than a 6- or 8-point one. Furthermore, it does not take more than twice as long to get a 31-point audiogram than an 8-point one. Our study’s results have two main disadvantages; these have to do with finding appropriate patients and studying high-frequency hearing loss. From the perspective of selecting patients, as explained in the methods section, there are many prerequisites for choosing subjects to participate. From the point of view of hearing loss, the processing did not give generate enough enhancements in hearing because of the nature of the subjects’ disparities in hearing loss and the use of the high-frequency loss spectrum. These effects can be seen from the aided thresholds (Fig. 1).
In the testing, the signal was presented to both ears. Thus, the experimental scheme may be suboptimal in cases where subjects have asymmetrical hearing thresholds at high frequencies. For the undirected sounds test, the increase of perceptual performance was very clear after two weeks. Although without being told the sound type, only 30.66% success was achieved in the first week as a mean, this value increased to 80.25% in the fifth week. Especially in terms of the four subjects who completed the study (subjects 9, 6, 8, 3), the improvements showed the reliability of the algorithm and the effects of the training (subject 9: from 33% to 83%; subject 6: from 20% to 83%; subject 8: from 13% to 77%; and for subject 3: from 47% to 78% as means). Undirected results are more important than the others because this type of input is closest to that which subjects encounter in their daily life. The biggest effect of the algorithm was demonstrated for noisy sounds, which is the one of the main problems for highfrequency subjects. Although a mean of only 25.33% success was achieved in the first week, this value increased to 87.5% in the fifth week for noisy sounds. Adaptation in the case of noise was very high for all subjects. Interesting values were found for subjects 7 and 9, because in the first week, subject 7 did not understand any of the sounds, but after one week of training, this rate increased to 52%; subject 9 increased his mean perceptual performance from 17% to 92% at the end of the fifth week. The worst performance was achieved in the environmental sounds test. Mean percentages occurred in the range of 35–68.75%. The highest individual improvement was shown in subject 6 (80%). The same trend was exhibited in the noisefree sounds test; only subject 3 exhibited a decline after two weeks and did not attain the first week’s percentage in the fifth week. Among the sound types (environmental, music, and speech), the highest mean perceptual performance was obtained for music sounds in the fifth week. Although the
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 2 ( 2 0 1 1 ) 277–287
mean percentage was high (56.44%) at the beginning of the study, it climbed to 95% by the end of the study. This shows a significant effect on understanding music sounds in our method. Another significant achievement was obtained in the speech sounds test after five weeks of training. Two subjects (subjects 8 and 3) exhibited a decline in the last two weeks, but overall success was enough to show the reliability of the application. In general, most of the mean percentages increased by 50% at the end of the study compared to the first week. This study was the one of the parts of a Ph.D. thesis and methods are now available only as software. The structure and complexity of the software were designed in thinking about the implementation standards for hearing aids and considering that implementation studies will soon be completed. Of course, that three subjects had dropped out at the end of the study and small number of participants are limitations of this study; however, as a preliminary result, some ideas can be inferred from the results obtained in the five weeks of the study. In future, a comparison with participants’ own hearing aids should be incorporated and the number of subjects should be increased.
Conflict of interest The authors claim no conflict of interest with regard to this work.
Acknowledgement The authors wish to thank Elif Isitme Ltd. for supplying the special test rooms for trials and organizing the subjects’ appointments, and especially Ms. Aynur Akdogan for providing subjects for their voluntary participation and supplying all needs in the study.
references
[1] P.G. Stelmachowicz, D. Lewis, S. Choi, B. Hoover, Effect of stimulus bandwidth on auditory skills in normal hearing and hearing-impaired children, Ear Hear. 28 (2007) 483–494. [2] P.G. Stelmachowicz, A. Pittman, B. Hoover, D. Lewis, Aided perception of /s/ and /z/ by hearing-impaired children, Ear Hear. 23 (2002) 316–324. [3] J. Elfenbein, M. Hardin-Jones, J. Davis, Oral communication skills of children who are hard of hearing, J. Speech Hear. Res. 37 (1994) 216–226. [4] J.C. Steinberg, M.B. Gardner, The dependence of hearing impairment on sound intensity, J. Acoust. Soc. Am. 9 (1937) 11–23. [5] T.Y. Ching, H. Dillon, D. Byrne, Speech recognition of hearing impaired listeners: predictions from audibility and the limited role of high-frequency amplification, J. Acoust. Soc. Am. 103 (1998) 1128–1140. [6] C.W. Turner, K.J. Cummings, Speech audibility for listeners with high-frequency hearing loss, Am. J. Audiol. 8 (1999) 47–56. [7] N.B. Murray, D. Byrne, Performance of hearing impaired and normal hearing listeners with various high frequency cut-offs in hearing aids, Aust. J. Audiol. 8 (1986) 21–28.
287
[8] J.A. Sullivan, C.S. Allsman, L.B. Nielsen, J.P. Mobley, Amplification for listeners with steeply sloping, high frequency hearing loss, Ear Hear. 13 (1992) 35–45. [9] T. Baer, B.C. Moore, K. Kluk, Effects of low pass filtering on the intelligibility of speech in noise for people with and without dead regions at high frequencies, J. Acoust. Soc. Am. 112 (2002) 1133–1144. [10] D.A. Vickers, B.C. Moore, T. Baer, Effects of low-pass filtering on the intelligibility of speech in quiet for people with and without dead regions at high frequencies, J. Acoust. Soc. Am. 110 (2000) 1164–1175. [11] B.M. MacArdle, C. West, J. Bradley, S. Worth, J. Mackenzie, S.C. Bellman, A study of the application of a frequency transposition hearing system in children, Br. J. Audiol. 35 (2001) 17–29. [12] H.J. McDermott, M.R. Knight, Preliminary results with the AVR ImpaCt frequency-transposing hearing aid, J. Am. Acad. Audiol. 12 (3) (2001) 121–127. [13] C.W. Turner, R.R. Hurtig, Proportional frequency compression of speech for listeners with sensorineural hearing loss, J. Acoust. Soc. Am. 106 (1999), pp. 887–886. [14] C.M. Reed, B.L. Hicks, L.D. Braida, N.I. Durlach, Discrimination of speech processed by low-pass filtering and pitch-invariant frequency lowering, J. Acoust. Soc. Am. 74 (1983) 409–419. [15] K. Yasu, M. Hishitani, T. Arai, Y. Murahara, Critical-band based frequency compression for digital hearing aids, Acoust. Sci. Tech. 25 (2004) 1. [16] T. Neary, Static, dynamic, and relational properties in vowel perception, J. Acoust. Soc. Am. 85 (1989) 2088–2113. [17] A. Simpson, A. Hersbach, H.J. McDermott, Improvements in speech perception with an experimental nonlinear frequency compression hearing device, Int. J. Audiol. 44 (2005) 281–292. [18] B. Johansson, A new coding amplifier system for the severely hard of hearing, in: Proceedings of the 3rd International Congress on Acoustics Stuttgart, 1961, pp. 655–657. [19] D. Ling, Three experiments on frequency transposition, Am. Ann. Deaf 113 (1968) 283–294. [20] M. Velmans, M. Marcuson, The acceptability of spectrum preserving and spectrum-destroying transposition to severely hearing-impaired listeners, Br. J. Audiol. 17 (1983) 17–26. [21] E. Wedenberg, Auditory training of the severely hard of hearing using a coding amplifier, in: Proceedings of the 3rd International Congress on Acoustics Stuttgart, 1961, pp. 658–660. [22] D.S. Bennett, V.W. Byers, Increased intelligibility in the hypacusic by slow-play frequency transposition, J. Aud. Res. 7 (1967) 107–118. [23] H.J. McDermott, V.P. Dorkos, M.R. Dean, T.Y. Ching, Improvements in speech perception with use of the AVR TranSonic frequency-transposing hearing aid, J. Speech Lang. Hear. Res. 42 (1999) 1323–1335. [24] W. Davis-Penn, M. Ross, Pediatric experiences with frequency transposing, Hear. Instrum. 44 (1993) 26–32. [25] T.C. Parent, R. Chmiel, J. Jerger, Comparison of performance with frequency transposition hearing aids and conventional hearing aids, J. Am. Acad. Audiol. 9 (1997) 67–77. [26] F. Kuk, P. Korhonen, H. Peeters, D. Keenan, A. Jessen, H. Andersen, Linear frequency transposition: extending the audibility of high-frequency information, Hear. Rev. 13 (10) (2006) 42–48. [27] J.D. Robinson, T. Baer, B.R.C. Moore, Using transposition to improve consonant discrimination and detection for listeners with severe high-frequency hearing loss, Int. J. Audiol. 46 (2007) 293–308. [28] The MathWorks, Inc., http://www.mathworks.com.
journal homepage: www.intl.elsevierhealth.com/journals/cmpb
Preliminary results of a novel enhancement method for high-frequency hearing loss Umut Arioz a,∗ , Kemal Arda b , Umit Tuncel c a b c
Medical Informatics, Informatics Institute, M.E.T.U, Turkey Radiology, Ankara Oncology Education and Research Hospital, Turkey Otorhinolaryngology, Ankara Oncology Education and Research Hospital, Turkey
a r t i c l e
i n f o
a b s t r a c t
Article history:
In this study, a software program was developed for high-frequency hearing loss subjects
Received 11 September 2009
that includes a detailed audiogram and novel enhancement methods. The software per-
Received in revised form
forms enhancements of the audibility of high-frequency sounds according to the subject’s
13 May 2010
detailed 31-point audiogram. This provides subject-specific gains in the entire frequency
Accepted 13 May 2010
spectrum, and especially for high frequencies, of sounds. Amplification, compression, and transposition are the three main processing methods used to obtain the desired enhance-
Keywords:
ments for the subjects. For low frequencies, only the amplification method was used
Amplification
according to the dB value of the input. For mid and high frequencies, the compression and
Compression
transposition methods were used together. To obtain the preliminary results of the study,
Transposition
10 subjects were enrolled in a detailed audiogram study for five weeks. In the study, envi-
Audiogram
ronmental, music, and speech sounds were used. While the perceptual mean performances
High frequency
of the subjects were in the range of 25.33–63.77% in the first week, those values increased
Hearing loss
to 68.75–95.75% in the fifth week. In particular, all noisy and speech sounds were more significantly identified and understood by the subjects with the enhancement method (from 25.33% to 87.5% and from 42.33% to 90.5%, respectively). Three subjects had dropped out at the end of the study and small number of participants are the limitations of this study; however, as a preliminary result, some ideas can be inferred from the results for a smaller set of subjects obtained in the five weeks of the study. Thus, the reliability of the study needs to be tested with more subjects and a comparison with their own hearing aids should be conducted. © 2010 Elsevier Ireland Ltd. All rights reserved.
1.
Introduction
Many hearing-impaired individuals have a greater loss of hearing sensitivity at high frequencies than at low frequencies. It has been shown that high-frequency acoustic cues impact speech-recognition abilities even for normal-hearing children [1]. Children with hearing loss are more affected by limited high frequencies in terms of speech recognition [2], and may
∗
also exhibit errors in morphology such as plural markers and verb tense [3]. Providing high-frequency audibility to subjects with severe high-frequency hearing loss presents several challenges for the clinician when fitting a hearing aid. Firstly, the level difference between discomfort and the threshold of audibility is usually much less than that of a normally hearing ear [4]. Secondly, acoustic feedback will often limit the amount of usable gain, depending on individual character-
Corresponding author at: Medical Education and Informatics, Hacettepe University, Sihhiye 06100, Ankara, Turkey. Tel.: +90 312 3052578. E-mail address: [email protected] (U. Arioz). 0169-2607/$ – see front matter © 2010 Elsevier Ireland Ltd. All rights reserved. doi:10.1016/j.cmpb.2010.05.004
278
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 2 ( 2 0 1 1 ) 277–287
istics such as how effectively the ear canal is sealed by the earmold. Up to now, researchers have reported the results of trying many different enhancement methods for high-frequency hearing losses in the literature. Amplification, compression, and transposition are the most frequently used techniques. However, all three methods have both advantages and disadvantages related to different parts of the frequency spectrum. Although conventional amplification can provide usable low-frequency information, the amplified high-frequency sounds of speech are often inaudible due to the severity and configuration of the hearing loss. Until recently, the findings have suggested that an increase in high-frequency gain may not improve, and in some cases may even degrade, speech recognition for listeners with high-frequency hearing losses [5,6]. Other investigators have shown some improvements in specific subjects (e.g., flat hearing loss) or under certain conditions (e.g., soft speech and fricatives) [7,8]. It is recommended that clinicians use caution in providing amplification in the high frequencies when the hearing loss in this region is greater than approximately 55 dB [9,10]. There is a need, therefore, to consider alternative sound-processing options for individuals with severe high-frequency losses. Ideally, this processing will preserve useful low-frequency amplification and provide some additional high-frequency information. An alternative approach might be the use of frequency compression or transposition schemes, whereby high-frequency signals are shifted to lower frequencies to provide adequate audibility. This type of approach has produced mixed results, with some studies showing substantial improvement and others showing no improvement or degradation in performance [11–13]. With frequency compression, high-frequency sounds are compressed into lower frequencies where there is better residual hearing and less amplification is required. This allows for a decrease in gain in the high frequencies where the subject’s hearing is more damaged. The decrease in gain will also reduce feedback, distortion, and discomfort. The lack of spectral overlap between the shifted and the unshifted signals can be identified as an advantage of frequency compression. On the other hand, that it does not preserve frequency ratios for high frequencies can be identified as a disadvantage of frequency compression. The perceptual performance of a number of linear and nonlinear frequency-compression schemes was evaluated in one study [14]. In a preliminary study, six subjects with normal hearing participated in experiments that investigated whether any of the schemes could improve the discriminability of consonant stimuli. Although none of the schemes provided better performance than a standard condition that applied only low-pass filtering to the stimuli, the best scheme was found to be a variant that progressively increased the amount of frequency compression for input frequencies above approximately 1200 Hz. In another study [15], the quality and intelligibility of speech sounds were improved for hearing-impaired subjects through compression along the frequency axis using an FFT-based approach. In this technique, the signal for each frame was transformed from the time domain to the fre-
quency domain using FFT. After the amplitude and phase spectra of the FFT were calculated, a have attained speech perception benefits when listening to proportional frequencycompression spectrum was computed for each band. Then, the processed spectrum was multiplied by the original phase spectrum to re-synthesize a band-limited signal. Finally, with the overlap add technique and IFFT process, the final signal was obtained by some listeners [13]. An advantage of this method is that the frequency ratios are preserved. These ratios may be particularly important cues for the recognition of vowels in speech [16]. In a recent study [17], nonlinear frequency compression was used with 1.6 kHz cutoff frequency and 0.5 compression exponent parameters. This research provided better recognition of monosyllabic words and achieved greater high-frequency sensation levels by enabling frequency compression than with conventionally fitted hearing aids. Various sound-processing schemes have been developed over the past decades that have attempted to present information from high-frequency regions of speech at lower frequencies. In extreme cases, frequency lowering (“frequency shifting” or “transposition”) may be the only way to provide this extra information acoustically. One scheme [18] implementing disproportionate frequency shifting, and several studies using early types of frequencylowering methods [19,20] reported little success in providing speech understanding benefits. This may have occurred for a variety of reasons. Much of the information related to the spectral shape of the incoming signal was lost as a result of the processing technique. These schemes may have provided additional high-frequency information at the expense of other perceptual cues by overlapping the shifted and unshifted signals. Some early attempts at frequency transposition converted signals in the 3–6 kHz range into low-frequency noise below 1.5 kHz by passing signals through a nonlinear modulator [18,21]. No significant improvements in speech intelligibility were found with the device [19,20]. Proportional frequency shifting, using a ‘slow play’ method, is an alternative sound-processing technique [22,23]. Segments of the speech signal are recorded and then played back at a slower speed than employed for recording. The TranSonic (AVR Sonovations Inc., USA, 1991) and ImpaCt DSR675 (AVR Sonovations Inc., USA, 1998) are hearing instruments that incorporate such processing. Incoming signals dominated by components at frequencies above 2.5 kHz are shifted down by a factor that is programmable for each listener. Positive outcomes were reported when the TranSonic was fitted to a small number of hearing-impaired children [24]. A similar study with adults demonstrated speech perception benefits with the device [23,25], and reported that the amplification characteristics of the TranSonic in the low frequencies, rather than its frequency shifting characteristics, may have provided most of the benefit. In another transposition study for children [11], the system categorized the incoming signal according to whether its energy peak was above or below 2.5 kHz. Signals below 2.5 kHz were categorized as vowels, and their transposition coefficients were in the range of 1–1.4; signals above 2.5 kHz were categorized as consonants, and their transposition coef-
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 2 ( 2 0 1 1 ) 277–287
ficients were in the range of 1–6. At the end of the study, only one-third of the initial group fitted was still using the device and obtaining improvements in the specific frequencies. In a more successful study [26], an audibility extender (AE) was employed as a frequency-lowering technique that uses linear frequency transposition to move inaudible highfrequency sounds to audible low-frequency regions. First, the algorithm receives information about the wearer’s hearing loss and decides which frequency region (source octave) will be transposed by the help of the AE’s signal processing component, which is called a dynamic integrator. The frequency at which transposition begins is called the start frequency. Typically, in which system, one octave of sounds above the start frequency was transposed. The linearly transposed signals were mixed with the original signal below the start frequency as the final output. As a result, subjects with a sloping highfrequency hearing loss subjectively preferred the AE for its much simpler stimulus. In a more recent study [27], an edge frequency (fe) parameter was selected as an edge frequency for the dead region. A band from 1 fe to 1.7 fe was selected as the destination for the transposed frequency components. High-frequency components from a source band well within the dead region (2–2.7 fe) were shifted into this band, but no frequency compression was applied. In the algorithm used, frequencies below fe were left unchanged. Transposition only occurred if the short-term spectrum was dominated by high frequencies. The results of the tests showed that although there was no significant overall benefit to transposition, the processing did not impair consonant identification. In our study, we used all three methods (amplification, compression, transposition) in an attempt to provide highfrequency audibility in high-frequency hearing loss patients. Because each of these methods has an advantage for a specific range of the spectrum, we planned to benefit from all of the techniques. In order to do so, software was developed to perform the signal processing automatically according to the patient’s audiogram. In this way, we wanted to overcome of the problem of generalization of the same algorithm for every patient which may be the main reason for unsuccessful results of the past studies. Thus, we developed a patient-specific algorithm based on detailed audiograms.
2.
Materials and methods
2.1.
Subjects
Ten hearing-impaired adults, two women and eight men, participated in the study. Detailed audiogram measurements had been taken from all subjects at the time of testing, but three subjects did not follow the study after one week. The majority of the subjects participated in the whole study for software sound testing (seven subjects for two weeks, five subjects for three weeks, and four subjects for five weeks). The number of subjects decreased because of personal reasons. The reasons of the subjects’ nonattendance were distance between their homes and test center and not having enough mobility to come the tests (subject 5); having huge hearing loss and not expecting a benefit from the study (subjects 1 and 10);
279
and having other health problems more important than hearing loss after the study began (subjects 2, 7, and 4). Relevant information about all the subjects is provided in Table 1. The majority of the subjects had moderate-to-severe hearing loss. Their hearing threshold levels, measured conventionally with headphones by the standard method, are listed in Table 2. Nine of the subjects were experienced hearing aid users. They had worn hearing aids for several years and wore their current hearing aids on a daily basis. The remaining subject had not previously worn a hearing aid (subject 5). The subjects were not paid for their participation in the experiment, although expenses such as travel costs were reimbursed. All explanations about the study were given to the subjects at the beginning of the study. The participation of the subjects in this experiment was in accordance with the ‘Guiding principles for research involving human or animal subjects’.
2.2.
Stimuli
To evaluate the efficacy of the enhancement algorithm, the aided hearing sensitivity, speech recognition, and identification of environmental sounds and music were measured to determine whether new high-frequency cues were available. Thus, seven sounds with different dB values from 20 to 80 dB range were prepared in the categories of environmental, music, and speech sounds with their noisy forms (a total of 19 for noiseless sounds [7 environmental sounds, 7 music sounds, and 5 speech sounds] and 17 noisy sounds [6 environmental sounds, 6 music sounds, and 5 speech sounds]). All sounds were selected according to their high-frequency content ([0–8 kHz] frequency spectrum range). Environmental sounds were selected by considering the high-frequency sounds in daily life that occur most frequently, like the sound of water in the kitchen, paper tearing, jingling keys, or two plates hitting one another. Music sounds were short selections from instrumental songs. Speech sounds were selected from the Turkish film cues, and incorporated different speaking properties. Noisy sounds were constructed by adding brown noise, pink noise, or white noise to the original sounds. This was done using the signal processing’s simple addition commands in the time domain. All noises were tried one by one to obtain the most suitable noisy sound and selected with the criterion of being not too dominant over the original sound. All noise sounds were taken from the standard sounds library of the MATLAB software [28]. For all subjects, determining and understanding the type of sound was conducted in five different ways (1. without saying the sound’s property or type [undirected] for all types of sounds; identifying the sounds as: 2. noisy sounds for all types of sounds, 3. environmental sounds for both noise-free and noisy sounds, 4. music sounds for both noise-free and noisy sounds, 5. speech sounds for both noise-free and noisy sounds).
2.3.
Software
All signal processing was done using our own software. The software was developed using MATLAB [28]. Mainly, the pro-
280
Left Binaural Right Right – Left Binaural Binaural Binaural Right Linear Linear Linear Linear – Linear Linear Linear Linear Linear Fully automatic digital Fully automatic digital Fully automatic digital Fully automatic digital – Fully automatic digital Fully automatic digital Fully automatic digital Fully automatic digital Digitally programmable Bernafon F 401 Bernafon Neo 401 Bernafon Win 112 Bernafon Win 401 – Bernafon Win 102 Audio service Nova 35 Bernafon Win 112 Audibel DeluxA Bernafon AF 310 Unknown Presbyacusis Industrial noise exposure High-level noise exposure Presbyacusis Presbyacusis Unknown Unknown Unknown Unknown 51 67 59 37 53 57 28 17 22 48 Subject 1 Subject 2 Subject 3 Subject 4 Subject 5 Subject 6 Subject 7 Subject 8 Subject 9 Subject 10
M M M F M M M M F M
Probable etiology of hearing loss Sex Age Subject
Table 1 – Subjects’ information about their hearing loss.
Type of own hearing aids
Features of own hearing aids
Processing strategy of conventional hearing device
Ears fitted
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 2 ( 2 0 1 1 ) 277–287
gram has two parts. In the first part, the detailed audiograms of the patients can be taken at 31 different frequency points. In the second part, the main signal processing can be done according to each patient’s detailed audiogram. Both parts have user friendly graphical interfaces. The user can construct any patient-specific test sounds using standard embedded sounds, his/her own specified sounds, or recorded sounds. The software works according to octave calculation of the sounds. By processing with octaves, the harmonics of the sound signal is preserved and distortions after processing are minimized.
2.4.
Signal processing
The frequency spectrum was split into three parts (0–1 kHz as low frequency, 1–4 kHz as mid frequency, and 4–8 kHz as high frequency) by taking into consideration the studies mentioned in Section 1. Thus, the goal was to avoid spectral or vocal distortions in the sounds throughout the whole process. Splitting into octaves is a critical process for applying the desired part of the spectrum with suitably without distorting its characteristic properties. Otherwise, unnatural sounds and overlaps and gaps occur; these are some of the main reasons for the unsuccessful results of earlier studies. The low-frequency part consisted of 3 octaves (125–250, 250–500, 500–1000); the midfrequency part consisted of 2 octaves (1000–2000, 2000–4000); and the high-frequency part consisted of 1 octave (4000–8000). Each octave’s value was determined as an average from its audiogram dB values. Audiogram frequency points were determined as giving enough information for low frequencies and detailed information for mid and high frequencies. Low-, mid, and high-frequency regions had the following 3 ranges of points that determined the enhancement parameters: 1. 2. 3. 4. 5. 6.
octave = 125–250 – 1 measurement – 1 range octave = 250–500 – 2 measurements – 1 range octave = 500–1000 – 4 measurements – 1 range octave = 1000–2000 – 4 measurements – 1 range octave = 2000–4000 – 8 measurements – 2 ranges octave = 4000–8000 – 12 measurements – 3 ranges
For the low-frequency part, only the amplification process was applied to the sounds. This application was conducted according to the octaves’ dB value. The gain was calculated by determining the decibel difference of the input signal and patient’s hearing loss for each octave. Thus, unneeded high amplification was not applied to the subjects (for example, for a 40 dB input sound, 10 dB amplification was applied for 50 dB hearing loss in that frequency range). For the mid- and high-frequency parts, frequency compression and transposition were applied as an enhancement. Frequency compression was applied for mid frequencies, while frequency transposition was applied for high frequencies. However, the algorithm worked by taking into account both the mid- and the high-frequency parts. The frequency compression and transposition parameters (compression region, compression rate, transposition region, transposition rate) were determined automatically by the software algorithm according to the audiogram of the patient. The criteria of the determination of the enhancement parameters were
281
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 2 ( 2 0 1 1 ) 277–287
Table 2 – Hearing thresholds of the subjects. Subject
Hearing loss (dB)/frequency (kHz) Ear
0.25
0.5
1
2
4
6
8
S1
L R
15 20
30 20
60 55
70 80
80 80
85 90
90 85
S2
L R
15 15
25 20
30 30
50 55
55 60
65 70
55 70
S3
L R
43 25
35 20
65 55
70 65
90 90
90 120
90 120
S4
L R
15 20
20 20
20 20
65 60
80 75
80 85
85 80
S5
L R
10 10
10 10
10 10
20 25
65 60
65 65
S6
L R
45 20
45 30
50 30
60 30
80 40
110 60
S7
L R
45 30
60 45
65 70
65 60
65 65
60 65
S8
L R
40 40
50 60
75 70
90 100
110 90
115 90
S9
L R
15 15
30 35
45 40
45 45
55 50
55 55
S10
L R
35 35
45 35
75 70
85 80
120 120
120 120
that no gap could occur and that there was no overlap between the compressed and the transposed parts. Thus, in a case where two processes (compression and transposition) occurred, the software searched for the compression ratio and transposition amount that were suitable to avoid overlapping or gap positions after the whole process had occurred. It did so by taking into account the shape of the audiogram. The aided thresholds for our software are shown in Fig. 1 for every point of four different patients’ audiograms. Enhancements for the mid- and high-frequency parts were applied according to the dB value of the input signal. After measuring the dB value of the input signal, no calculation was needed for any enhancement of the part of the frequency spectrum that was below the dB value, as this part was also audible for the patient. Enhancements were applied only to the part of the spectrum above dB value. Thus, fewer calculations and measurements were obtained for the same enhancement using a less time-consuming process. A flow chart of the enhancement processing can be seen in Fig. 2.
2.5.
Audiogram measurement
Detailed audiograms of all patients were taken by our software through the use of headphones. For the audiogram test, standard audiological test sounds (125, 250, 500, 1000, 2000, 4000, 6000, 8000 Hz) and detailed frequency sounds (750, 1250, 1500, 1750, 2250, 2500, 2750, 3000, 3250, 3500, 3750, 4250, 4500, 4750, 5000, 5250, 5500, 6500, 7000, 7500 Hz) were used that were reconstructed from standard sounds by sampling processing. All tests were done in a test room that was sound-isolated and specially designed for audiogram tests.
2.6.
Test procedure
Subjects were specially selected for our study. Because of our test procedures, there were some constraints involved that limited the number of subjects. At the beginning of the study, the requirements and obligations of the study were explained to the subjects; if they accepted them, the study was initiated. These requirements involved being able to come to test weekly, having time to listen to the sounds at least three times a day, and being able to listen the sound CD in their home with a computer or other electronic equipment with only headphones. The latter was made compulsory to obtain the same environment for all subjects. For the tests, a detailed audiogram was taken at the first meeting. Then, according to the subject’s audiogram, all sounds were enhanced with the software and stored on a CD. At the second meeting, a few days later, the prepared patient-specific sound CD and patient listening follow-up form were given to subjects; on the same day, the first test with enhanced sounds was conducted as a control study. After this, every 7 days, testing was done with the same enhanced patient-specific sounds and follow-up form was controlled.
3.
Results
3.1.
Detailed audiogram
The audiograms of the 10 subjects taken by both the standard method and our detailed 31-point method are shown in Fig. 3. The correlation coefficients between the audiograms were in the range of 0.9482–0.9889, with a mean of 0.9682. These were calculated after adding linearly the
282
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 2 ( 2 0 1 1 ) 277–287
Fig. 1 – Aided thresholds, example of four subjects (filled triangle: after enhancement, empty triangle: before enhancement).
missing values between the standard method values. The correlation coefficients of our audiograms and the standard measurement showed not only the reliability of the usage but also the need for more frequency points in the audiogram measurement. The difference and the additional frequency information obtained can be seen in Fig. 3. In particular, the difference and the extra frequency information can be seen in the 0–2000 Hz range for subjects 1, 2, 4, and 9; in the 2000–4000 Hz range for subjects 5 and 6; and in the 4000–8000 Hz range for subjects 4 and 6. For subjects 3 and 7, the 31-point audiogram did not provide extra information because of their flat and linear hearing loss configuration. Moreover, for subjects 8 and 10, the 31-point audiogram provided extra information only in the low frequencies because of the highest level of flat hearing loss in the high frequencies.
3.2.
Aided thresholds
All subjects obtained aided thresholds; as examples, four subjects’ aided thresholds, along with their pre-processing thresholds, are shown in Fig. 1. Subjects 1 and 2 obtained aided thresholds particularly in the 1500–3000 Hz range through the automatic addition of a 20–35 dB gain. Subjects 5 and 6 got aided thresholds in particularly the 2000–5000 Hz range through the automatic addition of 25–35 dB gain.
3.3.
Subject tests
In our study, all subjects listened to the sounds and gave responses like hearing/not hearing and understanding/not understanding. The percentages of both hearing and understanding responses were calculated for each subject every
Fig. 2 – Flowchart of the algorithm.
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 2 ( 2 0 1 1 ) 277–287
283
Fig. 3 – Detailed and standard audiograms of all ten subjects (filled circle: detailed audiogram values, empty circle: standard audiogram values).
284
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 2 ( 2 0 1 1 ) 277–287
Fig. 4 – Five weeks correct understanding percentage values of undirected sounds.
week. These values are shown in Figs. 4–8. Of all the subjects, seven came to second meeting; for the other three subjects; no comparisons were obtained for this study. Only 4 subjects completed all five weeks of the study. For all subjects participating in the study, the mean, minimum, and maximum values of the results are shown in Table 3. As shown, there were small decrements for noise-free sounds in the fourth week and for speech sounds in the fifth week; for the rest of the measures, an increase can be seen very clearly and significantly from the first week to the fifth week.
4.
Discussion
The main goal of our study was to offer patient-specific fittings for hearing losses. Differences between patients include
lifestyle, speech understanding, and cognitive ability. Each of these differences may result in one patient requiring a different fitting than another patient who has similar levels of hearing loss. In addition, any difference in the audiogram must be taken into account for better satisfaction and life quality. As seen in the literature, the amplification technique is much more efficient for low frequencies. Frequency compression can be considered in the mid frequencies, where amplification does not work or gives unsatisfactory results. Finally, frequency transposition can be used for frequency lowering of high frequencies to much more audible levels. By taking exploiting the advantages of the three main methods, a new method has been developed. In our method, amplification, compression, and transposition are responsible for the 0–1000 Hz, 1000–4000 Hz, and 4000–8000 Hz regions, respectively. These regions were determined by taking into
Fig. 5 – Five weeks correct understanding percentage values of noisy sounds.
285
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 2 ( 2 0 1 1 ) 277–287
Fig. 6 – Five weeks correct understanding percentage values of environment sounds.
Fig. 7 – Five weeks correct understanding percentage values of music sounds.
Table 3 – Mean, minimum and maximum values of perceptual performances of the subjects. First week mean (min–max) Undirected Directed Noiseless Noisy Env. sounds Music Speech
30.66 (13–49) 55.55 (27–81) 63.77 (33–89) 25.33 (0–45) 35 (0–56) 56.44 (40–75) 42.33 (6–78)
Second week mean (min–max) 42.28 (25–70) 75.42 (58–90) 74.14 (57–89) 50.28 (17–80) 49.85 (30–65) 70.85 (38–90) 66 (48–90)
Third week mean (min–max) 61.8 (50–73) 88.2 (68–100) 81.8 (67–92) 67.8 (46–88) 58.6 (30–75) 83.6 (50–100) 82.2 (75–89)
consideration of the sound octave parts and characteristics of the frequency regions. Thus, the distortion of the signals was ameliorated and the speech discrimination rates were improved.
Fourth week mean (min–max) 72.5 (67–78) 94 (83–100) 78 (63–92) 82 (75–93) 60.25 (38–80) 87.75 (85–90) 92.5 (85–100)
Fifth week mean (min–max) 80.25 (77–83) 95.75 (93–100) 82 (63–92) 87.5 (82–92) 68.75 (50–80) 95 (85–100) 90.5 (83–100)
Based on these observations, the audiogram configuration is suggested to be an important factor in recommending an enhancement algorithm. The frequency at which the hearing loss becomes severe, as well as the
286
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 2 ( 2 0 1 1 ) 277–287
Fig. 8 – Five weeks correct understanding percentage values of speech sounds.
steepness of the slope of the audiogram, may be important. The most important and vital role in determining the hearing aid algorithm or fitting is the patient’s audiogram, because most of the processes in this study were carried out according to it. Thus, true measurement is very crucial. Because of the non-objective nature of audiogram measurement, future research should seek to improve the accuracy of the measurement. In this study, we tried lessen the effect of non-objectivity by obtaining the measurement for more points relating to high-frequency loss in order to produce a more accurate audiogram measurement. In the measurement, standard audiogram methods were used. Using more points than are employed for standard measurement, not only was extra information gained that standard measurement cannot detect, but also there was no missing frequency information, which is especially important for high frequencies. Using more points in the audiogram measurement provided more accurate fitting algorithms for the patients; this results in more satisfaction in terms of patients’ hearing aids. This improvement, of course, is important and specific to high-frequency loss patients and our study’s methodology. In our approach, applying a more accurate fitting algorithm to patients is more important than the time it takes to obtain a 31-point audiogram rather than a 6- or 8-point one. Furthermore, it does not take more than twice as long to get a 31-point audiogram than an 8-point one. Our study’s results have two main disadvantages; these have to do with finding appropriate patients and studying high-frequency hearing loss. From the perspective of selecting patients, as explained in the methods section, there are many prerequisites for choosing subjects to participate. From the point of view of hearing loss, the processing did not give generate enough enhancements in hearing because of the nature of the subjects’ disparities in hearing loss and the use of the high-frequency loss spectrum. These effects can be seen from the aided thresholds (Fig. 1).
In the testing, the signal was presented to both ears. Thus, the experimental scheme may be suboptimal in cases where subjects have asymmetrical hearing thresholds at high frequencies. For the undirected sounds test, the increase of perceptual performance was very clear after two weeks. Although without being told the sound type, only 30.66% success was achieved in the first week as a mean, this value increased to 80.25% in the fifth week. Especially in terms of the four subjects who completed the study (subjects 9, 6, 8, 3), the improvements showed the reliability of the algorithm and the effects of the training (subject 9: from 33% to 83%; subject 6: from 20% to 83%; subject 8: from 13% to 77%; and for subject 3: from 47% to 78% as means). Undirected results are more important than the others because this type of input is closest to that which subjects encounter in their daily life. The biggest effect of the algorithm was demonstrated for noisy sounds, which is the one of the main problems for highfrequency subjects. Although a mean of only 25.33% success was achieved in the first week, this value increased to 87.5% in the fifth week for noisy sounds. Adaptation in the case of noise was very high for all subjects. Interesting values were found for subjects 7 and 9, because in the first week, subject 7 did not understand any of the sounds, but after one week of training, this rate increased to 52%; subject 9 increased his mean perceptual performance from 17% to 92% at the end of the fifth week. The worst performance was achieved in the environmental sounds test. Mean percentages occurred in the range of 35–68.75%. The highest individual improvement was shown in subject 6 (80%). The same trend was exhibited in the noisefree sounds test; only subject 3 exhibited a decline after two weeks and did not attain the first week’s percentage in the fifth week. Among the sound types (environmental, music, and speech), the highest mean perceptual performance was obtained for music sounds in the fifth week. Although the
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 2 ( 2 0 1 1 ) 277–287
mean percentage was high (56.44%) at the beginning of the study, it climbed to 95% by the end of the study. This shows a significant effect on understanding music sounds in our method. Another significant achievement was obtained in the speech sounds test after five weeks of training. Two subjects (subjects 8 and 3) exhibited a decline in the last two weeks, but overall success was enough to show the reliability of the application. In general, most of the mean percentages increased by 50% at the end of the study compared to the first week. This study was the one of the parts of a Ph.D. thesis and methods are now available only as software. The structure and complexity of the software were designed in thinking about the implementation standards for hearing aids and considering that implementation studies will soon be completed. Of course, that three subjects had dropped out at the end of the study and small number of participants are limitations of this study; however, as a preliminary result, some ideas can be inferred from the results obtained in the five weeks of the study. In future, a comparison with participants’ own hearing aids should be incorporated and the number of subjects should be increased.
Conflict of interest The authors claim no conflict of interest with regard to this work.
Acknowledgement The authors wish to thank Elif Isitme Ltd. for supplying the special test rooms for trials and organizing the subjects’ appointments, and especially Ms. Aynur Akdogan for providing subjects for their voluntary participation and supplying all needs in the study.
references
[1] P.G. Stelmachowicz, D. Lewis, S. Choi, B. Hoover, Effect of stimulus bandwidth on auditory skills in normal hearing and hearing-impaired children, Ear Hear. 28 (2007) 483–494. [2] P.G. Stelmachowicz, A. Pittman, B. Hoover, D. Lewis, Aided perception of /s/ and /z/ by hearing-impaired children, Ear Hear. 23 (2002) 316–324. [3] J. Elfenbein, M. Hardin-Jones, J. Davis, Oral communication skills of children who are hard of hearing, J. Speech Hear. Res. 37 (1994) 216–226. [4] J.C. Steinberg, M.B. Gardner, The dependence of hearing impairment on sound intensity, J. Acoust. Soc. Am. 9 (1937) 11–23. [5] T.Y. Ching, H. Dillon, D. Byrne, Speech recognition of hearing impaired listeners: predictions from audibility and the limited role of high-frequency amplification, J. Acoust. Soc. Am. 103 (1998) 1128–1140. [6] C.W. Turner, K.J. Cummings, Speech audibility for listeners with high-frequency hearing loss, Am. J. Audiol. 8 (1999) 47–56. [7] N.B. Murray, D. Byrne, Performance of hearing impaired and normal hearing listeners with various high frequency cut-offs in hearing aids, Aust. J. Audiol. 8 (1986) 21–28.
287
[8] J.A. Sullivan, C.S. Allsman, L.B. Nielsen, J.P. Mobley, Amplification for listeners with steeply sloping, high frequency hearing loss, Ear Hear. 13 (1992) 35–45. [9] T. Baer, B.C. Moore, K. Kluk, Effects of low pass filtering on the intelligibility of speech in noise for people with and without dead regions at high frequencies, J. Acoust. Soc. Am. 112 (2002) 1133–1144. [10] D.A. Vickers, B.C. Moore, T. Baer, Effects of low-pass filtering on the intelligibility of speech in quiet for people with and without dead regions at high frequencies, J. Acoust. Soc. Am. 110 (2000) 1164–1175. [11] B.M. MacArdle, C. West, J. Bradley, S. Worth, J. Mackenzie, S.C. Bellman, A study of the application of a frequency transposition hearing system in children, Br. J. Audiol. 35 (2001) 17–29. [12] H.J. McDermott, M.R. Knight, Preliminary results with the AVR ImpaCt frequency-transposing hearing aid, J. Am. Acad. Audiol. 12 (3) (2001) 121–127. [13] C.W. Turner, R.R. Hurtig, Proportional frequency compression of speech for listeners with sensorineural hearing loss, J. Acoust. Soc. Am. 106 (1999), pp. 887–886. [14] C.M. Reed, B.L. Hicks, L.D. Braida, N.I. Durlach, Discrimination of speech processed by low-pass filtering and pitch-invariant frequency lowering, J. Acoust. Soc. Am. 74 (1983) 409–419. [15] K. Yasu, M. Hishitani, T. Arai, Y. Murahara, Critical-band based frequency compression for digital hearing aids, Acoust. Sci. Tech. 25 (2004) 1. [16] T. Neary, Static, dynamic, and relational properties in vowel perception, J. Acoust. Soc. Am. 85 (1989) 2088–2113. [17] A. Simpson, A. Hersbach, H.J. McDermott, Improvements in speech perception with an experimental nonlinear frequency compression hearing device, Int. J. Audiol. 44 (2005) 281–292. [18] B. Johansson, A new coding amplifier system for the severely hard of hearing, in: Proceedings of the 3rd International Congress on Acoustics Stuttgart, 1961, pp. 655–657. [19] D. Ling, Three experiments on frequency transposition, Am. Ann. Deaf 113 (1968) 283–294. [20] M. Velmans, M. Marcuson, The acceptability of spectrum preserving and spectrum-destroying transposition to severely hearing-impaired listeners, Br. J. Audiol. 17 (1983) 17–26. [21] E. Wedenberg, Auditory training of the severely hard of hearing using a coding amplifier, in: Proceedings of the 3rd International Congress on Acoustics Stuttgart, 1961, pp. 658–660. [22] D.S. Bennett, V.W. Byers, Increased intelligibility in the hypacusic by slow-play frequency transposition, J. Aud. Res. 7 (1967) 107–118. [23] H.J. McDermott, V.P. Dorkos, M.R. Dean, T.Y. Ching, Improvements in speech perception with use of the AVR TranSonic frequency-transposing hearing aid, J. Speech Lang. Hear. Res. 42 (1999) 1323–1335. [24] W. Davis-Penn, M. Ross, Pediatric experiences with frequency transposing, Hear. Instrum. 44 (1993) 26–32. [25] T.C. Parent, R. Chmiel, J. Jerger, Comparison of performance with frequency transposition hearing aids and conventional hearing aids, J. Am. Acad. Audiol. 9 (1997) 67–77. [26] F. Kuk, P. Korhonen, H. Peeters, D. Keenan, A. Jessen, H. Andersen, Linear frequency transposition: extending the audibility of high-frequency information, Hear. Rev. 13 (10) (2006) 42–48. [27] J.D. Robinson, T. Baer, B.R.C. Moore, Using transposition to improve consonant discrimination and detection for listeners with severe high-frequency hearing loss, Int. J. Audiol. 46 (2007) 293–308. [28] The MathWorks, Inc., http://www.mathworks.com.