1. A segment of dialogue (3 secs.). 2. An excerpt from a Bach Brandenburg. Concerto (4 secs.). ..... schritte der Akustik â DAGA 93, Bad Honnef: DPG-GmbH, Teil ...
Quantification of significant sound quality attributes in the context of hearing instrument fine tuning Anne May, M.A. Claus Brenner Larsen, M.D. Asgaut Warland, M.D.
Michael Boretzki Phonak GmbH Fellbach-Oeffingen For a successful hearing instrument fitting, it is essential to meet three categories of auditory criteria: loudness/audibility, speech intelligibility and sound quality. Audibility and loudness are quantified by pure tone audiometry and loudness scaling. For measuring speech intelligibility a range of different procedures are available for both aided and unaided testing conditions. A method to quantify sound quality issues during the hearing instrument fitting is still being developed. Contributions to this issue have been made by Gabrielson et al. (1990) and Geers & Haubold (1993). The studies reported in this paper have been conducted by the research group of O. Heller, Würzburg, and are published for the first time in Boretzki et al. (1997a), Boretzki et al. (1997b) and Boretzki (1999).
25
Focus
News / Ideas / High Technology / Acoustics
During a fitting, the hearing impaired client reports to the clinician not only on audibility and loudness of certain sounds and the ability to understand other speakers, but also on sound quality e.g. the current prescription may render some sounds as “tinny”, “harsh” or “dull”. The clinician uses these descriptions to fine tune the hearing instrument to suit the client’s hearing impairment. For a psychoacoustic foundation to this approach three areas were investigated: 1. Dimensional analysis Which perceptual dimensions are affected by the different types of signal manipulation used in current hearing instrument technology e.g. spectral, dynamic changes, artifacts? 2. Psychophysical Analysis How do perceived sound changes correlate with the manipulation of the fitting parameters? 3. Analysis of the psychological type of sound quality attributes Are the sound quality attributes quantitatively stable, such as in “loudness”, and therefore easily investigated through a scaling procedure, or are we dealing with aspects of sound for which the client forms a frame of reference during the scaling procedure, which is then governed by the individually presented degradations of sound quality attributes? In the latter case, individual scalings of sound quality attributes could not then be compared easily to that of other subjects.
Study I: Analysis of Parameters The first step towards sound quality assessment is an analysis of the particular attributes of sound that are specific to hearing instrument fittings. This raises the question of whether hearing impaired people perceive sound quality on the same dimensions as those with normal hearing. The task is to procure precise descriptors for these attributes. Therefore the method used here is primarily based on the phenomena of sound and the descriptions used. Instead of laboratory generated sound, three different real-life sounds were used (see method “A-Life”, Geers & Haubold, 1993) which are sensitive to sound quality deficiencies: 1. A segment of dialogue (3 secs.). 2. An excerpt from a Bach Brandenburg Concerto (4 secs.). 3. A series of five chords played on a piano (6 secs.). These sound samples were systematically modified in a number of ways to create a range of samples representing different sound quality deficits. The manipulations covered most of the range of acoustical dimensions that are currently available in hearing instrument technology. The samples were also modified in a manner, so that normal hearing subjects were able to detect a difference between the original sound sample and it’s variant. Each modification was produced in a mild and a strong form, so that together with the original, a three-step change sequence was achieved. The subjects (13 with normal hearing thresholds and 8 with a hearing impairment) were asked to freely describe the perceptual changes in each three-step sequence.
3
News / Ideas / High Technology / Acoustics
The digitized change sequences were presented through loudspeakers as often as each subject wished in order to describe the perceived differences. There were five groups of variations for each of the three original sound samples: spectral and dynamic changes, level increase/ decrease, and the introduction of noise and distortion. The largest of these groups was made up of the 19 spectral changes. Nine variants representing a narrowband overemphasis were achieved by increasing the level by 10dB and 15dB in a one octave band with mid-frequencies in the range of 126–5011 Hz. Ten additional variants were achieved by bandpass filtering, each filter being three octaves wide with 9dB/octave and 3dB/octave filter slopes, and midfrequencies between 79 and 5011 Hz and represented the prominence of a larger frequency region. Table 1 summarizes the spectral variations. Five different types of compression (kneepoint at 58dB, below that squelch by expansion) and one expansion characteristic were used to represent modifications possible in non-linear instruments (attack time 1ms, three compression variants in two channels at a cross-over frequency of 1100Hz). Compression changes were introduced in the low frequency band, high frequency band, or both bands. The two single channel compression changes differed in that in one sequence the compression ratio was increased from 2:1 to 6:1; in the other, the attack time was reduced from 50 ms to 10 ms. The other parameters are listed in Table 2. The level changes included 4dB and 8dB increases and decreases from the original. To achieve a noisy signal, white noise of 40dB and 54dB was added to the original. Peak clipping was introduced to achieve the distortion samples.
4
Table 1 Spectral modifications An: One-octave wide level increases Bp: Three-octave wide bandpass filter strong
weak
third-octave bands from Hz
to Hz
center Hz
An0911+15
An0911+10
8.5-11.5
178
126
An1113+15
An1113+10
10.5-13.5
141
282
200
An1315+15
An1315+10
12.5-15.5
224
447
316
An1517+15
An1517+10
14.5-17.5
355
708
501
An1719+15
An1719+10
16.5-19.5
562
1122
794
An1921+15
An1921+10
18.5-21.5
891
1778
1259
An2123+15
An2123+10
20.5-23.5
1412
2818
1995
An2325+15
An2325+10
22.5-25.5
2239
4466
3162
An2527+15
An2527+10
24.5-27.5
3548
7079
5011
Bp0412 9/O+6
Bp0412 3/O+3
3.5-12.5
28
224
79
Bp0614 9/O+6
Bp0614 3/O+3
5.5-14.5
45
355
126
Bp0816 9/O+6
Bp0816 3/O+3
7.5-16.5
71
562
200
Bp1018 9/O+6
Bp1018 3/O+3
9.5-18.5
112
891
316
Bp1220 9/O+6
Bp1220 3/O+3
11.5-20.5
178
1412
501
Bp1422 9/O+6
Bp1422 3/O+3
13.5-22.5
282
2239
794
Bp1624 9/O+6
Bp1624 3/O+3
15.5-24.5
447
3548
1259
Bp1826 9/O+6
Bp1826 3/O+3
17.5-26.5
708
5623
1995
Bp2028 9/O+6
Bp2028 3/O+3
19.5-28.5
1122
8912
3162
Bp2230 9/O+6
Bp2230 3/O+3
21.5-30.5
1778
14124
5011
89
Table 2 Dynamic Variants KV: Compression ratio AZ: Recovery time (ms) TT: Low frequency channel HT: High frequency channel strong
weak
KV
TT-KV
TT-AZ
HT-KV
HT-AZ
Ex0.5-50
Ex0.75-50
0.5 / 0.75 50
AZ
-
-
-
-
K6-50
K2-50
6/2
50
-
-
-
-
K2-10
K2-50
2
10 / 50
-
-
-
-
KT3-10H3-10 KT3-50H3-50 -
-
3
10 / 50
3
10 / 50
KT3-10
KT3-50
-
-
3
10 / 50
-
-
KH3-10
KH3-50
-
-
-
-
3
10 / 50
Figure 1 Psychophysical profile and incidence of “dull” for each sample; 1a: normal hearing group, 1b: hearing impaired group. 1a
1b
The aim of the study was explained in detail to the test subjects. The complete procedure of freely describing the 87 sound samples (29 variants x 3 originals) took several hours of concentrated listening by each subject. The analysis of the descriptors used by more than one subject in the normal hearing group yielded 56 words. These fell into the following groups: loudness, hardness (hard, soft), shade (bright, light, dark), pitch (low, high), fullness (fuller, more volume, thinner), texture (tinny, woody, scratchy), voice quality (squeaky, nasal), articulation quality (hissy, breathy) time structure (clipped, reverberant,
echoing), distance (closer, further away), clarity (unclear, blurry), pleasantness of sound (unpleasant, ugly, bad). The hearing impaired subject’s descriptions varied from the normal hearing group in these respects: 1. They did not detect the difference in some of the change sequences, especially in the narrow band increases of the high frequencies. This was expected due to the fact that a hearing impairment was present. 2. For low frequency emphasis, most of the hearing impaired group reported that annoying background noises had vanished. 3. Some of the high frequency variants were clearer to the hearing impaired than the original sound samples, also expected due to the hearing loss. 4. The impaired group, again expectedly, described the variants of the dialogue sample more in terms of speech intelligibility. 5. Some of the variants were perceived as being more pleasant sounding (more intelligible, warmer). The normal hearing group detected deterioration in nearly all of the variants in comparison to the original samples. However, none of the differences between the two groups indicate a difference of dimensional structure. They simply show different levels on the same range of dimensions, with a different importance rating. Therefore it is possible to use normal hearing subjects for the purpose of further investigation of a scaling procedure for sound quality. From a methodological point of view, normal hearers have the advantage of being a homogeneous group.
5
Fig.1 shows the incidence of the term “dull” for both listener groups. The example demonstrates that the psychophysical profiles for both groups are comparable. The low frequency 3-octave wide bandpass filters were more often described as “dull”, whereas the 1-octave level increases in the low frequencies were deemed “dull” less often. Fig. 2 shows a further example for the less often cited term “tinny”. The incidence profiles of the 56 descriptor terms were analysed with respect to their sensitivity for particular acoustical modifications. The result is an inventory of 21 descriptors (sometimes two words) indicating special transmission deficits: volume (unspecific), dull, damped (low frequencies), dark (low frequencies), bright, harsh (high frequencies), shrill (narrow band high frequency emphasis), sharp (high frequencies), hard (more likely high frequencies), voluminous, full (low frequencies), thin, flat (more likely high frequencies), hollow (narrow band over amplification), echoing, reverberation (compressed), intermittent, (expanded), tinny, metallic (high frequencies), boomy (low frequencies), nasal (more likely high frequencies), vibrating (distortion), scratchy, crackling (distortion), hissing (high frequency compression), unclear, blurry (unspecific), unpleasant (unspecific).
6
Figure 2 Psychophysical profile and incidence of “tinny” for each sample (normal hearing subjects).
Study II: Sound feature scaling The starting point for studies II and III was the inventory of the 21 sound features indicating acoustical modifications. The aim was to develop an inventory of descriptors which the hearing impaired client can use during the fitting process to report on the sound of the hearing instrument quantitatively, by way of a categorical scale. For this purpose, two scaling experiments were initiated to test the following: 1. Whether the approach of using a categorical scale for the description of sound features is, in principle, promising. 2. How precise categorical descriptions of sound quality are, by means of an inventory of sound features. 3. Whether the psychophysical profiles, which were derived from the qualitative study could be reproduced in the quantitative descriptions and 4. If the sound quantification by way of an inventory is stable when the set of samples to be described is shifted from an equally distributed one to a set of mainly high frequency emphasized variants.
Figure 3 Rotation wheel for quantitative sound feature description
In both studies the same sound material was used as in the qualitative study, except that the variants consisted of only the strong modifications of each sample. Each presentation was preceded by the original sound sample. In experiment III a more reduced variation series was utilized. In experiment II, 26 normal hearing subjects (Bach: 8, dialogue: 9, piano: 9) described each of the variants (repeated as often as necessary) by using the inventory of 21 features, on the rotation wheel (Fig. 3). The subject’s task was to establish whether the feature shown at that time on the rotation wheel accurately described the sound sample, and if so, how strongly. The advantage of the rotation wheel was that the subject was forced to consider each of the 21 attributes in turn on an equal basis. The wheel showed
more than one attribute at a time so that subjects reported on 2–5 attributes at a time. If one of the attributes applied, then the subject reported the strength on a scale of 2 categories (“somewhat” and “fairly”. Fig. 3). In structure and distinction, the scale is based on the free description of sound features by normal and impaired listeners. The categories “somewhat” and “fairly” were deemed to be equally spaced according to an initial pilot study. The first feature “loudness” was, in each case, described by a category partitioning scale (Heller, 1985), which had 5 categories with ten steps each. Figs. 4 and 5 show the psychophysical profiles of the 21 features, which could be derived from the ratings. The modification sensitivities, which were attained in the qualitative study (Study I), were mirrored by the quantitative descriptions (Study II). Subgroups of the 21 features showed similar patterns (eg. dull, damped, dark), so that the inventory could be reduced further without significant loss of information. It is also clear that the same modification does not produce equal perceptual ratings across the 3 sound samples. For example, only the piano sample was perceived as strongly “echoing” or “reverberant”, especially the compressed variants. It is also clear that one specific modification is rarely recognized as showing only one feature (exception: only expansion led to the perception of “intermittency”). In summary, most signal modifications showed changes in more than one feature.
7
The standard deviations, for all attributes except “loudness”, are about a third of the scale’s range at the middle of the scale where the largest spread occurs. For “loudness”, the standard deviation is in the order of 10% of the scale. The spread is influenced by clear individual differences in the scaling which were present at the start of the rating session and need to be addressed further. For some of the dimensions, individual’s interpretations of descriptors play a part in the variability.
Figure 4 Psychophysical profiles, Study II, Part 1.
Study III: Sound feature scaling Study III showed that the large amount of spread in the scales was not related to either a low level of resolution in the scale or to the possibility of too great a demand on the listeners due to the large inventory (21 descriptors). Only a reduced set of modifications to the piano sample had to be described three times, for only 8 features. This study was also conducted with normal hearing subjects. Group A (8 subjects) was asked to use a 5-category scale (slightly, somewhat, moderately, fairly, very) each with 3 numerical steps. They did not show a reduced variability in the ratings compared to the control group (2 categories, 8 subjects). The spread was, as in Study II, a third of the scale. Using three runs of the description of each variant allowed discrimination between intra- and inter-individual variation. About half of the overall variation is due to intra-individual variance. This is independent of feature dimension and experimental group. Group B (8 subjects) used the same twocategory scale as the control group to describe but the sample series included mainly variants with high frequency emphasis. The spectral variants for the control group and group A formed an equally distributed series with respect to prevalence of a certain frequency region. This series of mostly high tone sound samples was used to check whether 8
the ratings were changed in comparison to the control group, in the presence of this unbalanced series. In practice, this is comparable to a fitting process that starts with too much high frequency gain and reduces this in small steps until the optimum fitting is reached. The fact that the subjects
had to describe more than one sound feature was intended to reduce the effect of context. During successive scalings of only one feature, the subjects would tend to develop a frame of reference for the set of presented stimuli. The results indicate that there is, in fact, no shift in the scaling due to a particular series of variants. Fig. 6 compares the psychophysical profiles of the three groups. The results show that the approach of using an inventory of descriptors, is a practically viable method to quantify sound quality perceptions, despite a marked scattering in the scale. The large spread of the ratings and the primary effect* show that, at the beginning of the rating session it is not clear to the subject what is a weak or strong example of a certain feature i.e. there is no stable reference system compared to loudness scaling. It was not until the first run of the study, that the subjects were able to orientate themselves with the sound modifications. Therefore the very early ratings already show large inter-individual variation. This variability remains through the entire course of the experiment (primary effect). The scaling methodology was shown to be robust in relation to context variation and yielded clearly defined psychophysical profiles, which, in summary, show themselves to be of practical value in the hearing instrument fitting process. Certainly the method described here lends itself to any further investigation into the psychophysical basis of modifying sound features for the hearing instrument fine tuning process.
Figure 5 Psychophysical profiles, Study II, Part 2.
9
Figure 6 Psychophysical profiles, Study III.
10
Further interesting questions for investigation are: Do the psychophysical reference points of the sound features adapt with slowly progressive hearing loss? Clinical experience shows that high frequency impairment does not result in everything sounding “dull”; recruitment does not cause all sounds to be “chopped” or “intermittent”, such as normal hearing subjects would describe high frequency attenuated or expanded sound samples. Hearing impaired people have a wide range of sound descriptors such as “dull”, “bright” or “shrill” available to them for description during a hearing instrument fitting. Determining precisely how severely each sound dimension is lacking is restricted by the fact that each person has to develop a
reference system in relation to each quality during the scaling process or the fitting time. However, the ability to determine whether a sound deficit is present or not is a reliable tool to deal with such a deficit. The psychophysical profiles are able to point to which fitting parameter is best used for the purpose of rectifying the particular sound quality deficit. * Primary effect: the rating of the first sample influences subsequent ratings.
Bibliography Boretzki M., Knoblach W., Pietzsch T., Haubold J., Heller O. (1997a) Analyse der psychologischen Dimensionen von hörgerätetechnischen Signalmanipulationen – Oder: Kann ein Hörgerät näseln? In: Wille P. (Hrsg.), Fortschritte der Akustik – DAGA 97, Oldenburg: DEGA, S. 85–86. Boretzki M., Fichtl E., Knoblach W., Pietzsch T., Haubold J., Heller O. (1997b), Skalierung phänomenaler Eigenschaften von hörgerätetechnischen Signalmanipulationen – Oder: Wie blechern klingt das Hörgerät? In: Wille P. (Hrsg.), Fortschritte der Akustik – DAGA 97, Oldenburg: DEGA, S. 87–88. Boretzki M. (1999) Quantifizierung bedeutsamer Klangeigenschaften im Hinblick auf die Hörgeräteanpassung. Zeitschrift für Audiologie, Supplementum II, 166–172. Gabrielsson A., Hagerman B., Bech-Kristensen T., Lundberg G. (1990) Perceived sound quality of reproductions with different frequency responses and sound levels. JASA 88, 1359–1366.
Michael Boretzki Ph. D Audiologist with Phonak Germany Michael Boretzki studied psychology at Würzburg University and gained ten years experience in the research group of O. Heller. His main areas of interest were perceptual, psychoacoustic and audiological research and education. He developed a number of software tools for perceptual experimentation, e.g. loudness scaling. His doctoral thesis covers questions in the area of psychophysical scaling. Since 1997 he has worked at Phonak, mostly on the development and application of hearing aid fitting methods. Thanks to Agnes Opp-Enzinger and Peter Hofmann, who conducted the experiments.
Geers W., Haubold J. (1993) Hörgeräteanpassung mit natürlichen Klangbildern. In: Fortschritte der Akustik – DAGA 93, Bad Honnef: DPG-GmbH, Teil B, S. 756–759. Heller O. (1985) Hörfeldaudiometrie mit dem Verfahren der Kategorienunterteilung (KU). Psychologische Beiträge 27, 509–519.
11
28603(GB)/0600 • na/Focus 25_GB/0600 • Printed in Switzerland © Phonak AG All rights reserved
www.phonak.com
