• WANTED: Happy members who like to discuss audio and other topics related to our interest. Desire to learn and share knowledge of science required. There are many reviews of audio hardware and expert members to help answer your questions. Click here to have your audio equipment measured for free!

Digit Triplet Test: Hear Speech in Noise

pozz

Слава Україні
Forum Donor
Editor
Joined
May 21, 2019
Messages
4,036
Likes
6,827
The test: https://www.axdesign.co.uk/tools-and-devices/digit-triplet-test
The background research: https://journals.sagepub.com/doi/10.1177/2331216519858301

My results are attached. I used AKG K371s and my laptop's terrible headphone output.

You'll notice one empty result. I didn't hear jack when the SNR was -10dB. Yes, negative 10dB. Meaning the noise was that much higher than speech.

Otherwise, few mistakes here and there. I have normal audiometric hearing. Not perfect, just normal, meaning all of my results on my audiograms are within 20dB HL.

Try it out yourself.
 

Attachments

  • results.pdf
    11 KB · Views: 137
OP
pozz

pozz

Слава Україні
Forum Donor
Editor
Joined
May 21, 2019
Messages
4,036
Likes
6,827
I've gotten a PM about this test expressing some confusion and irritation at the procedure.

First things first. I had no involvement in it. I sit on a listserv of academics that work in fields related to hearing and got an email about it.

In terms of the test itself, the design is skilled. Double blind and precise. But the interface kind of sucks and there's little explanation.

It involves 24 procedurally randomized trials. In each trial a British lady recites three numbers to you with noise in the background. The purpose is to understand how well people understand speech under noisy conditions. So the starting SNR is 0. In other words, speech and noise are at the same level. As you progress, which means correctly identifying all three spoken numbers in a trial, the SNR gets worse. What's worse than 0 is negative, which means the noise is higher in level than the speech. That's done either by increasing the level of the noise or decreasing the level of the speech. SNR, if you're not clear about what that is, is signal to noise ratio, the relative difference in level between some signal (speech) and noise. I'm not sure if there was a spectral component to the noise in the test which varied based on how well you're doing. Maybe, but most likely any differences are a function of level and nothing else.

There are number keys and red and green buttons. Green is to submit an answer. Red is to correct a mistyped figure.

All 24 trials took me under 10 minutes to finish. I went pretty fast, and on top of the test noise I could hear my dishwasher and washing machine running in the background. Not ideal conditions to get the best results, but whatever.

After you finish you get that autogenerated report I posted above. I think the columns are self explanatory apart from the "Reversal" column. What it indicates is the number of mistakes I made, with the test taking a step back and making the next trial easier. It works like the Klippel distortion test, if you've ever taken it (although the Klippel test shows far greater awareness of the test taker's experience and good administrative design).

If you're curious about why you can hear speech buried in noise, it's because noise is a probabilistic signal, with semirandom energy distributed across a certain bandwidth. Because of that, there are holes in it in which other signals, like speech, can reside. Since speech signals represent correlated (if complex) energy, and we're used to hearing those patterns, we can pick them out.

Poorer results generally mean hearing damage or something genetic. But there is also training and vocation to consider. Anyone whose job involves active listening will tend to do better.
 

dougi

Addicted to Fun and Learning
Joined
May 31, 2020
Messages
845
Likes
763
Location
ACT, Australia
Interesting. It would be good to get the results of the study eventually. I am 53, with tinnitus, hearing tested a couple of months ago with about 40dB loss at 8kHz but considered OK for my age. My result was -5.07dB (corrected from 5.7) S/N. What to take from it? Does this mean that some system noise contributions are not that important, or that you should aim for noise contributions lower than the room noise?
 
Last edited:

dougi

Addicted to Fun and Learning
Joined
May 31, 2020
Messages
845
Likes
763
Location
ACT, Australia
I was very interested in this, as I am 70 y.o. and have pretty bad hearing. My result is -4.52 dB. One odd thing is that my test had 28 trials, not 24. I have no idea why.

Very nice little tool. Thank you! Jim Taylor
Mine had 26 (0-25). It probably repeats one at the same SNR if you get one/some wrong, for repeatability.
 

dougi

Addicted to Fun and Learning
Joined
May 31, 2020
Messages
845
Likes
763
Location
ACT, Australia
The test: https://www.axdesign.co.uk/tools-and-devices/digit-triplet-test
The background research: https://journals.sagepub.com/doi/10.1177/2331216519858301

My results are attached. I used AKG K371s and my laptop's terrible headphone output.

You'll notice one empty result. I didn't hear jack when the SNR was -10dB. Yes, negative 10dB. Meaning the noise was that much higher than speech.

Otherwise, few mistakes here and there. I have normal audiometric hearing. Not perfect, just normal, meaning all of my results on my audiograms are within 20dB HL.

Try it out yourself.
LOL, mine didn't even try to do a test at -10dB. Mine was with laptop and Bose QC25ii with NC on.
 

Blumlein 88

Grand Contributor
Forum Donor
Joined
Feb 23, 2016
Messages
20,524
Likes
37,057
I ended up with -5.43 db for whatever it is worth.

I did something similar as a volunteer in college. They were saying numbers embedded in noise plus whispered low level words being repeated in the midst of the noise. You were interested in the numbers. You'd have noise, and low level words like, spider, spider, spider or ruler, ruler, ruler repeated while during this time numbers were spoken.
 
  • Like
Reactions: trl

JRS

Major Contributor
Joined
Sep 22, 2021
Messages
1,158
Likes
1,003
Location
Albuquerque, NM USA
I have terrible hearing, though afraid to test it. Scored -6.8 and were it not for my new Sundara's I would have scored much, much worse, maybe even >0. Now if only hearing aids could be made this good....

I read the accompanying article to understand the rationale--a most worthwhile project--I mean really 3,000. bucks for some mid-fi, cosmetically neutral one size fits all w/r to FR, to say nothing of all the artefact, dreck. They really could be sold DTC and make very good money. Do a test like we did measuring FR info and intelligebility info online, send the user some thermoset plugs to push in and make ear canal models, and pay your 750.00 for something really helpful. I'd do it in a heartbeat. And I would happily suffer some aesthetics for good battery life (which could be extended by having some clip on inductors for my reading glasses and I'ze be fat.

Thanks for posting--this is a big issue that doesn't get much press except for the occasional Miracle Ear stuff.
 

Attachments

  • results.pdf
    19.7 KB · Views: 72

JRS

Major Contributor
Joined
Sep 22, 2021
Messages
1,158
Likes
1,003
Location
Albuquerque, NM USA
Ah, yes. That would explain a thing or two. Jim
Yep. Sort of like having your eyes checked, they likely go back and forth until you can't reliably here a thing. Often I'd miss the first number and get the second and third, likely because I was cued to start listening carefully. She had clear enunciation in an impeccable accent, be curious if instead it were some dude in Mumbai.
 

JSmith

Master Contributor
Joined
Feb 8, 2021
Messages
5,153
Likes
13,219
Location
Algol Perseus
SNR: -6.47db
I didn't hear jack when the SNR was -10dB.
Seems I picked more correct at that level... yet still got a worse score. :)

20 2 -10 2,1,9 2,1,9
28 5 -10 3,6,0 3,6,0
36 8 -10 3,6,6 3,6,6
37 8 -10 4,6,0 4,6,0

Also, this test went right up to 40 attempts... so I gather it's averaging the results.

This can all depend on the numbers spoken... some are easier to hear over the noise than others due to the nature of the spoken sound.

Further, some may have issues with the accent if they're not English. No issue for me, but then again if there was that Aussie twang present I may have picked more... hard to say.

What would also help is if one could replay the sample again before entering the numbers.

I just did this on cheap earbuds on my laptop... will try again later on my main system.



JSmith
 

Attachments

  • results.pdf
    20.5 KB · Views: 74

MRC01

Major Contributor
Joined
Feb 5, 2019
Messages
3,424
Likes
4,030
Location
Pacific Northwest
... Further, some may have issues with the accent if they're not English. No issue for me, but then again if there was that Aussie twang present I may have picked more... hard to say.
...
I too had difficulty with her accent; her 0 and 4 sound similar to me. I only realized that after reviewing my first test results, I made a few errors with 0 and 4, but not with any other digits. PS: score -6.74.

Is there any information on score percentiles versus age? I didn't see this on the page or background research. Without this, the score doesn't really say much.
 
Last edited:

charleski

Major Contributor
Joined
Dec 15, 2019
Messages
1,098
Likes
2,239
Location
Manchester UK
SNR -5.22dB

There are some possible methodological problems here, though.
Firstly, some digits are more distinguishable simply because of the sounds they incorporate. A sibilant like 'six' is more easily identified than softer words like 'oh', 'one' or 'four', which are more readily confused.
Secondly, repeating digits within a phrase allows the use of alternate strategies. I noticed at least one example where I wasn't entirely sure of the first digit, but the second was the same sound and I was able to identify the digit by comparison since they came so close together, thus essentially averaging out some of the noise.
I also thought the opening phrase 'the digits ...' seemed to vary slightly in level between tests. Maybe they're looking at some sort of priming effect here. Or maybe this was just illusory because of the changing levels of noise.
 

tomtoo

Major Contributor
Joined
Nov 20, 2019
Messages
3,607
Likes
4,514
Location
Germany
-3.x the first, 0.0 the second run. Looks like all the tests are the same. You have to adopt to them. ;)
With this test, lay back, smoke some weed or sip a half glass of redwine and you get better. Espeacelly if your mouther thoung is not english.
 
Last edited:

MRC01

Major Contributor
Joined
Feb 5, 2019
Messages
3,424
Likes
4,030
Location
Pacific Northwest
I think lower scores are better: the score represents the level of the voice to the noise. Like, -3 dB means you can discern the voice when it's 3 dB below the level of noise.
 

Lambda

Major Contributor
Joined
Mar 22, 2020
Messages
1,785
Likes
1,519
I did it with my PC Speaker (5"KRK Rokit active monitor) at relatively low levels (since its night hear and peapole are sleeping in the next room...)
Anyways got -6.5dB.

Also they never said 7

EDIT:
Did it again with Headphones and got -7.76dB

Sadly i miss typed a number toads the end and i'm to lazy to do it again...
 
Last edited:

Killingbeans

Major Contributor
Joined
Oct 23, 2018
Messages
4,089
Likes
7,551
Location
Bjerringbro, Denmark.
-6.86dB with suboptimal listening conditions. Fun test :)

Looking at the results, I only misheard one specific digit; '0'. I heard it as '4' five times and marked it as '9' once.
 
OP
pozz

pozz

Слава Україні
Forum Donor
Editor
Joined
May 21, 2019
Messages
4,036
Likes
6,827
Interesting. It would be good to get the results of the study eventually. I am 53, with tinnitus, hearing tested a couple of months ago with about 40dB loss at 8kHz but considered OK for my age. My result was -5.07dB (corrected from 5.7) S/N. What to take from it? Does this mean that some system noise contributions are not that important, or that you should aim for noise contributions lower than the room noise?
Both conclusions work. But the thing to notice is how much effort noisy circumstances require of us. Reducing system noise or room noise will make listening easier and more precise.

I'm 33 and have fairly serious tinnitus.
 
OP
pozz

pozz

Слава Україні
Forum Donor
Editor
Joined
May 21, 2019
Messages
4,036
Likes
6,827
SNR -5.22dB

There are some possible methodological problems here, though.
Firstly, some digits are more distinguishable simply because of the sounds they incorporate. A sibilant like 'six' is more easily identified than softer words like 'oh', 'one' or 'four', which are more readily confused.
Secondly, repeating digits within a phrase allows the use of alternate strategies. I noticed at least one example where I wasn't entirely sure of the first digit, but the second was the same sound and I was able to identify the digit by comparison since they came so close together, thus essentially averaging out some of the noise.
I also thought the opening phrase 'the digits ...' seemed to vary slightly in level between tests. Maybe they're looking at some sort of priming effect here. Or maybe this was just illusory because of the changing levels of noise.
We don't know all the mechanics of the test. The envelopes and spectra of different consonants are understood and I would expect the design to take them into account.

There could be patterns to do with speech/noise levels that attempt to minimize the relative difficulty of each triplet.
 
OP
pozz

pozz

Слава Україні
Forum Donor
Editor
Joined
May 21, 2019
Messages
4,036
Likes
6,827
Is there any information on score percentiles versus age? I didn't see this on the page or background research. Without this, the score doesn't really say much.
This is a normal audiometric test.

Here's a random example:
1634946959759.png

Scatter plot of the digit triplet SRTs (only test, no retest) versus the average pure-tone thresholds at 500, 1000, 2000, and 4000 Hz. Crosses (x) represent normal-hearing (NH) ears; squares ( ) represent hearing-impaired (HI) ears. Arrows indicate data points with invalid SRT due to an incorrect response on the maximal SNR ( 8 dB). The horizontal lines represent the borders between a ‘good’, ‘insufficient’, and ‘poor’ outcome for the digit triplet screening test by telephone.

Different languages also have different results:
1634947286722.png


I think we shouldn't take the absolute numbers too seriously. Test circumstances dictate a lot.
 

tomtoo

Major Contributor
Joined
Nov 20, 2019
Messages
3,607
Likes
4,514
Location
Germany
I think lower scores are better: the score represents the level of the voice to the noise. Like, -3 dB means you can discern the voice when it's 3 dB below the level of noise.

Iam relative sure i did better in my second test.
Different language, and i needet some time to get used. It would be realy interesting what it realy says the db number. Hard for me to imagine i did the second test worse
 
Last edited:
Top Bottom