Beyond speech intelligibility

Authors
Thomas Lunner* †
Eliane Ng†
Maria Grube Sorgenfrei §
Marianna Vatti*
Renskje Hietkamp*
Dragan Gusatovic§
Graham Naylor*
*Eriksholm Research Centre,
Denmark
www.eriksholm.com
§Oticon A/S, Smørum, Denmark
ratios is difficult due to ceiling effects. Here we
suggest using working memory as an outcome
measure in high SNR conditions. See also the
companion poster Ng et al. (2014).
Contact
Thomas Lunner
[email protected]
Participants
125
250
500
Results
Hz
750 1000 1500 2000 3000 4000 6000 8000
100
0
tTwenty-five native Danish speakers (11 women and 14
men)
10
80
20
tSymmetrical moderate to moderately-severe, acquired
sensorineural hearing loss (see left-right average
audiograms in the figure)
30
40
tRecruited from the Eriksholm test subject pool
tAverage age 70 years (SD = 7.7, range: 54-79 years)
tAverage pure-tone thresholds (PTA) at .5, 1, 2 and 4 kHz in
both ears was 49.7 dB HL (SD = 9.9 dB HL)
Background
Noise reduction schemes for hearing aids have their own ‘ecology’ in the sense that they
should be tailored to work best where they are needed (and are able to deliver some noise
reduction). This calls for investigations that focus on daily life environments where benefit
of noise reduction can be shown.
tNo history of otological problems or psychological disorders
was reported.
Cognitive Spare Capacity (CSC) refers to the residual capacity after successful speech
perception. In a recent study by Ng et al. (2013), they defined the residual capacity to be
number of words recalled after successful listening to a number of HINT sentences, that is
the memory for heard sentences in ca 95% correct speech intelligibility. Ng et al. (in press)
conducted their test in the Swedish language, and showed that a noise reduction scheme
inspired by Wang et al. (2009) and Boldt et al. (2008) improved memory recall.
Changing the language for a test method may not always be straightforward. Our lab has
experience with some strange effects occurring when seemingly only changing language
(Lunner et al, 2012). Therefore we thought that a replication of the Ng et al. (in press) study
was motivated.
70
0
-10
-5
SNR (dB)
90
100
0
5
15
10
Average SNR = 9.6 dB (SD = 2.3)
Individually estimated SRT95%
Memory recall test
1.
2.
3.
4.
5.
6.
7.
Working memory (WM) is important for online language processing in a dialogue. We use
WM to store, to inhibit or ignore what is not relevant, and to attend to things selectively.
According to the Ease of Language Understanding (ELU) model (Rönnberg et al., 2008,
2013), a poor representation of a signal in the neural pathways will lead to activations of
the (effortful) WM system. We argue that hearing impaired persons must rely much more
on effortful WM resources to understand what has been said, compared to normal hearing
persons who can rely more on effortless highly over-learned automatic speech recognition
systems. Therefore tests which probe the degree of load on WM resources may be useful for
evaluating the benefit of hearing aids in terms of effort rather than performance.
Benefit of hearing aid signal processing is often assessed by speech intelligibility in noise
tests. Usually such tests are most sensitive at a signal-to-noise ratio (SNR) below 0 dB.
However, a recent study by Smeds et al. (2012) showed that the SNRs in ecological listening
situations (e.g. kitchen, babble, and car) were typically well above 0 dB SNR. That is, SNRs
where the speech intelligibility in noise tests are insensitive. Therefore new measures are
needed that can show eventual benefits of hearing aid processing in the + 5-15 dB range.
40
20
80
tAll were hearing aid users.
60
50
60
Her går alle med solbriller
Han ligger stadig i sengen
Eleven skriver en lang rapport
Hele byen kom til brylluppet
Hans datter vil på højskole
I går havde filmen premiere
Fabrikkens port var ikke lukket
*
***
**
NR OFF
NR ON
n.s.
The task was to repeat the last word after each sentence followed by recalling as many words as possible
after sentence #7 (Method inspired by Pichora-Fuller, 2006 and Sarampalis et al., 2009).
Procedure
1-2 (Primacy)
Memory recall at SRT95%
Stimuli:
3-5 (Asymptote)
6-7 (Recency)
Memory recall with serial positions.
tPrimacy -> Long-term memory
tDanish HINT sentences
tRecency in short-term memory (working memory)
tFour-talker babble (Danish)
Discussion & conclusion
tPresented via insert earphones
SNR for 95 % correct on average 9.6 dB (Range 6 - 15 dB)
Prepare Experiment:
√ In line with Ecological SNRs according to Smeds et al. (2012)
Assure audibility = No Processing condition (NoP). Linear gain with individually fitted frequency response
(linearized VAC rationale) to assure audibility up to 6.5 kHz.
Possible to show improved recall for an experimental NR algorithm at ‘high’ SNRs
√ Main effect of Aggressive noise reduction
Adjust SNR to ~95 % correct HINT sentences for NoP in 4 talker babble (SRT 84% + 2.2 dB)
√ Replicated Ng et al. (in press) in new language = robust effects
Proof of concept Memory test
Long-term memory improved by Aggressive noise reduction
Research question and hypothesis
Q: Is the memory effect found in Ng et al. (in press) robust in a new language (Danish)?
H: Aggressive noise reduction will improve memory recall relative to (individualized) linear
amplification in the Danish memory test, when tested in babble noise at positive SNRs
resembling a daily life situation.
Contrast: NoP vs. Aggressive NR in 4 talker babble, 5 repetitions
√ Indicates better neural encoding
Aggressive noise reduction (NR)
Aggressive noise reduction tries to enhance the talker and suppress competing noise/talkers (Boldt et al., 2008).
Limitations
- only tested with one speech material (HINT)
†Linköping University, Sweden
www.headcentre.se
Scan to access
poster pdf
Audiogram
% Correct
performance at high ecological signal-to-noise
Methods
dB HL
Beyond speech
intelligibility
testing:
A memory test
for assessment of
signal processing
interventions
in ecologically
valid listening
situations
Hearing aid assessment of speech intelligibility
- we do not know if 7 items recall is optimal for all subjects
Storage
Person in quiet
Memory Score = 6
…same person in noise
Memory Score = 4
…and even more noise
Memory Score = 2
…effect of HA in
even more noise
Memory Score = 4
Inspired by Pichora-Fuller (2006)
Implications
Increased evidence for a useful method to evaluate noise reduction systems at ecological SNRs
1
if s(t,f) - n(t,f) > LC
0
otherwise,
IBM(t,f) =
[LC = 0 dB]
REFERENCES
Boldt et al. (2008). Estimation of the ideal binary mask using directional systems. http://www.iwaenc.org/proceedings/2008/contents/papers/9062.pdf; Lunner et al. (2012) Effect of Speech Material on the Benefit of Temporal Fine Structure
Information in Speech for Young Normal-Hearing and Older Hearing-Impaired Participants. Ear and Hearing; Ng et al . (2013). Effects of noise and working memory capacity on memory processing of speech for hearing-aid users. Int J Audiol; Ng et
al., (in press). Noise reduction improves memory for target language speech in competing native but not in foreign language speech. Ear and Hearing; Ng et al. (2014) Noise reduction improves memory for target speech heard in competing speech.
POster at IHCON 20014.; Pichora-Fuller (2006) Perceptual Effort and Apparent Cognitive Decline: Implications for Audiologic Rehabilitation. Seminars in Hearing; Rönnberg et al. (2008). Cognition counts: A working memory system for ease of
language understanding (ELU). Int J Audiol; Rönnberg et al. (2013). The Ease of Language Understanding (ELU) model: theoretical, empirical, and clinical advances. Front Syst Neurosci; Sarampralis et al. (2009) Objective measures of listening effort:
effects of background noise and noise reduction. J Speech Lang Hear Res; Smeds et al. (2012) Estimation of Realistic Signal-to-Noise Ratios. Poster presented at IHCON 2012; Wang et al. (2009) Speech intelligibility in background noise with ideal
binary time-frequency masking. J Acoust Soc Am.