7 Keystones of Accurate HiFi Blind Test [Article]

Audio Basis - educational articles

"Audiophile myths debunked" is one of popular topics at a forums and an articles.
As rule, these "myths" is based on listening impressions, unexplained technically or small measurable difference is inaudible theoretically / probably.
To accurate check the "myths" recommended to use method, ensuring repeatability with same results in the similar conditions.

HiFi blind test is way of "objective" measurement of "subjective" perception.

Read about most important things, turning a HiFi test into a safe evidence.

7 Keystones of HiFi Blind Test [Article]

HiFi blind test is technique when listener (participant) try recognize unknown sample, sounding during check.

Sample here is record, or equipment, or software, or apparatus/software mode.

This trial is attempt to measure non-measurable music feature - perceived sound quality.

Objective here is measurable and repeatable in equal conditions.

Subjective here is detected by human feelings.

Still we cannot direct access to foreign feelings. We cannot listen music as other person.

We can ask him/her only. But they can't exactly describe their feeling. Especially in the subtlest details, providing by modern musical equipment, software, HiFi test records.

But we want to know: it is real audible or not.

It is main aim of the blind trial. We try eliminate subjective part of perception (price, bias, habit, etc., i.e. probable imaginary perception).

For even more "subjectivity" reducing, double blind test is used.

Double blind test is a trial where neither listener nor conductor know what is sample.

Blind test audio

Blind test audio

ABX test may be used in ear comparison.

ABX test is trial where samples A and B may be compared with sample X (either A or B). Listener should recognize what is X (A or B).

ABX test audio

ABX test audio

ABX audio test software

Foobar2000 with ABX test software plugin may be used to ABX test of audio files. However author don't know what happens with audio stuff when ABX-test-foobar's plugin prepare files before comparison (progress when the plugin start).

For ABX test Mac OS software you can try find ABXer utility. For iPhone exists ABX Tester application. Also available cross platform Lacinato ABX. Author don't learned these pieces of software, so he has no opinion.

If somebody claim that sound is different, it can't measure its audibility. Because we listen in the brains via ears.

Theoretically, measurement tools are more sensitive than human ears. But without trials we don't know where is audibility edge of difference.

Audibility edge is:

maximal value of a feature or
difference of values between listening samples,

that cannot be distinguished by human.

Professional double blind test of music equipment is not home entertainment. It is hard long expensive work.

It doesn't necessarily audiophile blind test should be done at laboratory. At home nobody can prohibit us to do it.

But safe trial results are necessary as claim evidences.

Proper blind Hi-Fi test features

These features are necessary to provide safe trial results:

Methodology
Protocol
Big number of measurements
Listening place
Testing equipment issues
Measurement precision issue
Careful condition control

Test should ensure repeatability in the same conditions.

Methodology

Trial begins with methodology design. Methodology define trial's:

aim,
precision,
implementation,
other things, noted below.

Methodology should be designed first

Methodology hi-fi blind test audio

Test protocol

Test protocol is paper, where recorded detailed data about conducted trial:

equipment,
measurement tools,
participants (and its listening skill, occupation, education, other),
trial conditions,
other.

Hi-Fi test protocol

Hi-Fi test protocol

Testing and measurement equipment may be registered with unique identifier of item (serial number, as example).

The main protocol aim is ability to check experiment conditions in case of doubt or to better understand result reasons.

Listening place

To save time several participants may be accommodated in one room. In the room are placed several seat rows.

In the trial result interpretation, necessary take into account:

each seat impact to sound,
seat with and without listener impact differently,
speaker sound is different for each seat.

Speaker have individual radiation pattern.
Frequency response depend on listener place relatively speakers and listening room's walls. Because acoustical rays are interfere, bounce from a surfaces and interfere too.

Impact seat place to audio test results

Impact seat place to audio test results

Anechoic room allow to avoid bouncing. But speaker have different frequency response in the different directions. So different seat places also cause different frequency response in anechoic room even.

Generally, only 1 sitting place in the listening room is recommended.

Testing equipment issues

Testing equipment (apparatus that will tested in the trial) should be checked to workability by way, that may be described in the methodology of the trial.

It is desirable, if testing equipment have unique identifier (serial number). It help to repeat experiment or found difference result reasons in case of doubt.
Because different equipment instances may have various figures and sound differently.

Very important thing!

LOUDNESS OF COMPARED SAMPLES

Human perception edge is close to 1 ... 2 dB of loudness. So 0.1 ... 0.2 dB level normalization of samples is recommended.

Human have limited time of echoic memory (auditory event retaining) [2].
So time sample listening should not be too long. Immediate (real-time) easy switch between samples must be ensured.

Updated: When we compare headphones, tested sample can't be isolated from participant technically. For solving this issue, synthesis of headphone features and comparison via single headphone unit were suggested [3].

Measurement precision issue

In statistical calculations, measurement error values may be different. We don't know it exactly, but accept in first approach, that it have normal distribution [1]. So...

To provide measurement precision X, it is recommended measure value Y with precision X/3 or higher.

As example, we want to provide ear trial measurement precision 1%.

1/1% = 1/0.01 = 100 tryings.

So trying number 3 times higher 300 = 100 * 3 is recommended.

Defined measurement precision should be provided

Measurement precision

Big number of measurements

The ear test have a many variables due human factor, above all.

Somewhat, it may be compensated by big number of tryings, participants, equipment items.

Big participant and trying number is recommended

Big participant and trying number

In the "Measurement precision issue" part was considered example: providing 1% precision.

There are recommended 300 tryings.

However, it don't guarantee 100 % of true.

Participant skills can cause biasing of results, as example.

To compensate it with precision 1%, we should invite 300 = (1/1%)*3 = (1/0.01)*3 participants.

Thus total number of measurements is 90000 = 300 trying * 300 participants

We should account each feature, that can cause mistake probably. To a first approximation, we should invite participants with different skills and groups of participants with same skill should have same number of members.

In the trial conclusions we should group results by participant skills.

General rule:

number measurements is enough,
if adding of several participants and/or tryings and/or equipment items and/or other
impact to results into allowable error margin.

Careful condition control

Conditions define experiment. Changing conditions can bias results significantly.

Careful control of conditions provide the test exactness

Careful control of coditions provide the test exactness

Example:

Somebody want to compare DACs. He asks peoples who have the devices, check it on some public HiFi test CD.

These tests, as rule, performed at home without exact condition control. We have experiment at different speakers, listening places, the ambient noise, etc. Even experiment, described in details, may have subtle, at first glance, issues, that impact upon result.

As rule, before experiment we can't know exactly, that details are important in the trial. Only result figures can show, that details are unimportant in our experiment: impact to result into allowable error margin.
As example, we cannot expect listening skill affected trial results. But for some tests, it might be important.

Conclusion

Blind test, double blind test, ABX test are intended to "objective" measurement of "subjective" (immeasurable) perception.
Properly performed and recorded hi-fi blind test may be a serious objective evidence of subjective perception.
The trial may be performed at home for personal purposes, but can't be used as technical evidence.
The trial should be designed, performed in a manner that maximizes independence from number of participants and tryings and/or other. As reference value may be allowable changes into error margin.
The independence is achieved by big number of participants and tryings, equipment items, etc.
Even correctly designed and performed HiFi test don't guarantee absolute truth. Because new knowledge may open new details.

Share the article in the social networks, forums

    Write comments, your thoughts about hifi blind tests
        at the article post on Audiophile Inventory's facebook page
        and elsewhere.