Can the use of subgroups in music mixing improve the preference of a mix and what perceived qualities are most prevalent in preferred mixes, as well as mixes with and without subgroups?

(1)

Can the use of subgroups in music mixing improve the preference of a mix and what

perceived qualities are most prevalent in preferred mixes, as well as mixes with and

without subgroups?

Gustav Björkman

Audio Technology, bachelor's level 2021

Luleå University of Technology

Department of Social Sciences, Technology and Arts

(2)

Abstract

Subgrouping is a mixing technique that most mix engineers use. Little research on perceptual evaluations of mixing techniques has been done and what little research that has, has been done on automatic mixing systems. When it comes to research on

subgrouping in automatic mixing systems the results show a significant preference towards automatic mixes with subgroups over mixes without. This study aims to test the same notion that the use of subgroups improves listener preference but this time in mixes created by humans. A group of mix engineers created two mixes of one song, one with the use of subgroups and one without. These mixes were the stimuli of a listening experiment that was conducted to investigate listener preference of mixes with and without subgroups as well as what perceived qualities were most prevalent in preferred mixes and mixes with and without subgroups. The results showed that mixes without subgroups were preferred over mixes with subgroups, although, these results were not statistically significant. The results also showed that balance, frequency and clarity were the most prevalent sonic qualities that helped the listeners decide how to rank the mixes. 

(3)

Abstract 1

Acknowledgements 3

1. Introduction 4

1.1. Background 4

1.2. Research question 6

1.3. Aims and purpose 6

2. Method 7

2.1. Creating stimuli 7

2.2. Listening experiment 7

2.3. Demographic data 7

2.4. Equipment and control room 7

2.5. Data analysis 8

3. Results and analysis 9

3.1. Preference 9

3.2. Perceived qualities 11

4. Discussion 14

5. References 16

6. Appendix 17

6.1. Instructions (A) for the mix engineers 17 6.2. Instructions (B) for the mix engineers 19 6.3. Instructions for the listening experiment 21

(4)

Acknowledgements

A big thank you to Jon Allan my supervisor who provided me with the right guidance and questions to help me get to the finish line. Thank you to all mix engineers who created mixes for this study, without you this study would not exist. A bunch of hugs and thank yous to my friends and classmates who gave me the right support when the going got tough. 

(5)

1. Introduction

Subgrouping is a mixing technique that, according to Ronan, Gunes & Reiss (2017), most engineers utilise when mixing music. Subgrouping involves outputting several channels to a separate bus before going to the master-bus.

This makes it very easy to apply the same changes to all subgroup members, like for example global signal processing and level adjustments. There is also hierarchical subgrouping which involves outputting several subgroups or buses to another bus before going to the master bus. Little research has been done on this topic and what little research that has, has been done in the purpose of developing intelligent audio production tools. However, given the prevalence of subgrouping, a deeper understanding of how subgrouping impacts a mix could be beneficial to human engineers too.

1.1. Background

Ronan et al. (2017) surveyed ten award winning mix engineers in a study about subgrouping techniques among mix engineers. The survey consisted of 21 questions, for which most of them were based on 9 assumptions that the authors had made. Some example of the assumptions are: Mix engineers create subgroups within subgroups (Hierarchical subgrouping), mix engineers subgroup to achieve subgroup effect processing and mix engineers subgroup based on instrument family. The authors found that there are many reasons mix engineers subgroup and that the most common reasons are the following:

- to maintain good gain structure

- to achieve subgroup effects processing

- to create individual submixes and to make the mix process less complicated They also found that mix engineers subgroup the same or similar instruments together. Furthermore they base their subgrouping decisions on the genre of music being mixed, they also apply hierarchical subgrouping on drums, guitars and vocals and lastly they base the amount of subgroups on the amount of tracks in the session.

Ronan, De Man, Gunes & Reiss (2015a) researched the perceptual impacts that subgrouping has on music mixes. The authors examined a dataset of mix

sessions compiled for an experiment by De Man, Boerum, Leonard, King, Massenburg & Reiss (2015). Ronan et al. (2015a) extracted data manually from these sessions. The following data was extracted: how many subgroups there was in each session, if EQ and compression was applied on the

subgroups, if there were any subgroup sends (aux send), what instruments there were in each subgroup and whether the subgroup was hierarchical.

Ronan et al. (2015a) calculated an overall preference score for each engineer

(6)

drums and vocals were hierarchically subgrouped without any other instrument types. The most hierarchically subgrouped instrument was drums. Hierarchical subgroups were also present in 19 of the 72 mixes. The results show moderate to strong correlations between subgroup to audio track ratio per mix engineer, with and without processing and mix preference rating. The results also show weak to moderate correlations between subgroup to audio track ratio per mix, with and without processing and mix preference rating. According to Ronan et al. (2015a) this shows that the correlation between the amount of subgroups and high mix preference when looking at mixes individually is very weak, but very strong when looking at mix engineers individually.

These two papers are of course closely linked, especially because they are written by some of the same authors but there are certainly differences in subject demography and method. Ronan et al. (2017) focuses on professional, award winning mix engineers while the subjects in the paper by Ronan et al.

(2015a) are in majority audio engineering students. Ronan et al. (2017) conducted an online survey questionnaire in order to collect the data while Ronan et al. (2015a) simply extracted relevant data from an, already existing, dataset. This means that the topic has been researched with two different methods, one that has given information provided by expert mixers as data and one that has given perceptual ranking as data. Both papers conclude that drums are hierarchically subgrouped the most and that vocals are

hierarchically subgrouped second most.

Research into automatic mixing systems further supports the notion that subgrouping leads to preferred mixes. Ronan, Ma, Namara, Gunes & Reiss (2018) tested the three following hypotheses:

- Is our proposed automatic mixing system able to be used to reduce the amount of auditory masking that occurs in a multitrack mix and

subsequently improve its perceived quality?

- Is using subgroups when generating an automatic mix able to improve the perceived quality and clarity of a mix?

- Is the use of subgroups in an automatic mixing system able to have an impact on the perceived emotions of the listener over automatic mixes that do not use subgroups?

The authors generated two automatic mixes for 5 multitrack sessions from The Open Multitrack Testbed (De Man, Mora-Mcginity, Fazekas & Reiss 2014). One of the mixes was generated using subgroups and the other mix was generated without subgroups. The mixes with subgroups were created using the

automatic subgrouping method that was tested in previous research (Ronan, Gunes, Moffat & Reiss 2015b) One more mix was created as a simple sum of the tracks as they were in the multitrack session, and two additional human mixes were created as well. The authors conducted two listening experiments with two parts to each where the second part was common to both. The first

(7)

part of one listening experiment (E1) asked the subjects to rate the five mixes of each song in terms of their preference. The first part of the other listening experiment (E2) asked the subjects to rate the five mixes of each song in terms of how well they could distinguish each of the sources present in the mix. The second part of both experiments (E3) asked the subjects to rate the two automatic mixes of each song for their perceived emotion along three scales. The scales were Arousal, Valence and Tension (A-V-T).

The authors found that there was a definitive preference towards the automatic mixes with subgroups over the automatic mixes without subgroups. The mixes created by humans had the highest preference of all for every song. The

authors found that the automatic mixes using subgroups had a higher

preference of clarity over the automatic mixes without subgroups. The authors found that their third hypothesis was true in only 1 of 15 cases (5 songs

measured along 3 affect dimensions).

These results show that subgrouping does impact mixes and more often than not, in a positive way. In both E1 and E3 the results show that subgrouping in automatic mixes increases perceived quality and clarity in the mix. It can be concluded that subgrouping does not help evoke emotional responses in

listeners, at least not in automatic mixes. The relevant findings in this research are not necessarily generalisable to human mixes and could use further

research to test the first two hypotheses (and perhaps also the third hypothesis) with mixes created by humans.

1.2. Research question

The research question is based on one of the hypotheses that Ronan et al.

(2018) tested and is formulated: Can the use of subgroups in music mixing improve the preference of a mix and what perceived qualities are most prevalent in preferred mixes, as well as mixes with and without subgroups?

1.3. Aims and purpose

The aim with this study is to test the preference of mixes with subgroups against mixes without subgroups. Furthermore this study aims to investigate what perceived qualities are most prevalent in, both preferred mixes and mixes with and without the use of subgroups. The purpose of the study is to gain more understanding of what subgrouping does for a mix, beyond allowing for subgroup processing and aiding the mix workflow.

(8)

2. Method

2.1. Creating stimuli

The first part of the study consisted of generating the stimuli for the listening experiment. Five professional mix engineers created two mixes of one song each. One mix where they utilised subgrouping and one mix where they didn’t.

The engineers were provided with the multitrack and an instruction document (Appendix 1 & 2). Each engineer worked in the control room where they

usually work, using the DAW that they feel the most comfortable with. Every engineer was instructed to set the pan-law of their DAW to -3 dB, in order for the differences of the mixes to be solely based on how each engineer mixes.

Three of the engineers started with Mix A (without subgroups) and the other two started with Mix B (with subgroups) in order to combat cognitive bias. The engineers were limited to the use of two plugins, Melda production’s

MCompressor and MEqualizer. One more mix was created as a simple sum of the audio tracks in the session to be used as an anchor in the listening

experiment. Every mix was loudness normalised to -23 LUFS using the normalisation tool in Logic Pro X. The song that was mixed is called Sea of Leaves by Jokers, Jacks and Kings and was retrieved from The open multitrack testbed (De Man et al. 2014).

2.2. Listening experiment

The second part of the study was the listening experiment. Seventeen

experienced listeners (audio engineer students and music students) took part in a listening experiment. A modified MUSHRA test was used. In each trial of the MUSHRA, the subjects ranked a 30 second snippet of each mix from one engineer, as well as the anchor, on a scale from 0 to 100 in order of

preference. The subjects also left a motivation for each mix as to why they ranked it the way they did. There were five trials of the test, one trial for each mix engineer.

2.3. Demographic data

The subjects of the listening experiment were 5 women and 12 men. The most preferred genre of music between the subjects was Pop with 4 subjects

followed by Indie with 3 subjects. The median age between the subjects was 23.

2.4. Equipment and control room

The listening experiment took place in the acoustically treated control room called K2 at Luleå Tekniska Universitet in Piteå. The experiment was conducted using a PC laptop with the software STEP, developed by Audio Research Labs.

(9)

The audio interface that was used was an RME Babyface connected to a pair of Klein & Hummel O410 monitors.

2.5. Data analysis

Since there was no given reference for “preference” in the listening

experiment, subjects used their own interpretations and thus the data had to be normalised using min-max normalisation. The function that was used was as follows: where y = normalised data, x = original data, a = the minimum datapoint and c = the range of the dataset. The quantitative data from the listening experiment was analysed with a paired t-test in order to see if the results are statistically significant. The qualitative data from the listening test was analysed with thematic analysis in order to find sonic qualities that correlate with high and low mix preference as well as mixes with and without subgroups.

y = (x − a)( 100 c )

(10)

3. Results and analysis

Two subjects’ set of quantitative data had to be discarded due to them not following the instructions correctly. One subjects set of qualitative data had to be discarded due to the same reason. This means that the data that was not discarded from these subjects (qualitative and quantitative, respectively) could still be used.

3.1. Preference

In Figure 1 the spread of data before being normalised can be seen and in Figure 2 the spread of data after being normalised can be seen. All further graphs and calculations are using the normalised data.

Figure 1: The spread of data for each trial of the listening experiment before min-max normalisation.

(11)

The data for Mix A, Mix B as well as the anchor was combined and is presented in Figure 3.

Figure 2: The spread of data for each trial of the listening experiment after min-max normalisation.

Figure 3: The combined spread of data for Mix A, Mix B and the anchor respectively

(12)

The results of the listening experiment show that Mix A (M = 78.16, SD = 20.02) tended to be ranked higher than Mix B (M = 72.93, SD = 24.09). There is however no statistical significance that the use of subgroups was a factor, t(74) = 1.56, p = 0.12.

3.2. Perceived qualities

The motivations of mix ranking from the subjects of the listening experiment were analysed using thematic analysis which resulted in themes, either

positive or negative. Mentions of each theme were counted for the highest and second highest mixes, as well as for Mix A and Mix B. The amount of negative mentions was subtracted from the amount of positive mentions to see the difference between positive and negative mentions. The themes and

explanations of the themes can be seen in Figure 4.

Figure 4: The themes from the thematic analysis as well as explanations for each theme.

Theme Explanation

Balance The balance of levels and instruments in the mix.

Frequency The frequency contents of the mix.

Clarity How well the listener can hear individual instruments.

Stereowidth The stereo width of the mix.

Drums The instrument drums.

Vocals The instrument vocals.

Dynamics The dynamics and compression of the mix.

Guitar The instrument guitar.

Bass The instrument bass.

(13)

Figure 5: The amount of mentions of each theme, positive, negative and the difference, for the highest ranked mixes.

The most prevalent themes from the highest ranked mixes are balance,

frequency and clarity, the highest being balance with 29 positive mentions and 6 negative mentions, as seen in Figure 5.

(14)

The most prevalent themes from the second highest ranked mixes are balance, frequency and clarity, the highest being frequency with 2 positive mentions and 31 negative mentions as seen in Figure 6.

Figure 5 and 6 show that the subjects of the listening experiment tended to talk more positively about the higher rated mixes than the second highest rated mixes. This is to be expected when asking subjects to rank and motivate their ranking. One outstanding theme within the second highest rated mixes is frequency with -29 as difference. This, together with a difference of 15

frequency mentions from the highest rated mixes, shows that frequency played a large part in the subjects ranking decisions. Clarity was the most positively mentioned theme for the second highest rated mixes with a difference of 11, although not nearly as positively mentioned as the highest rated mixes. This could say something about the multitrack’s recording quality being very high since clarity was positively mentioned in both the highest rated mixes as well as the second highest rated mixes.

Figure 7: The amount of mentions of each theme, positive, negative and the difference, for Mix A.

The most prevalent themes from Mix A are balance, frequency and clarity, the highest being balance with 22 positive mentions and 7 negative mentions as seen in Figure 7.

The most prevalent themes from Mix B are balance, frequency and drums, the highest being balance with 14 positive and negative mentions respectively as seen in Figure 8.

(15)

Figure 8: The amount of mentions of each theme, positive, negative and the difference, for Mix B.

One outstanding theme within Mix B is clarity which is the most positively mentioned theme. This could support the notion that the use of subgroups increases clarity that Ronan et. al. (2018) found to be true about the use of subgroups in automatic mixes. This could also be because of the before mentioned recording quality of the multitrack since clarity has the same difference, 12, in both Mix A and B. Frequency is another outstanding theme within Mix B since its difference is -2 while it is -7 for Mix A. This could mean that being able to use equalisation on more than only the individual tracks in the session improves the frequency spectrum. The mentions of balance differs a lot between Mix A and Mix B where Mix A has a difference of 15 and Mix B has a difference of 0. This could be because more compression is present since it could be applied to subgroups as well as individual channels. More

compression of course makes the material less dynamic which could lessen the perceived balance of a mix.

4. Discussion

The expected outcome of this study was to find that mixes with subgroups are ranked higher than mixes without. This was not the case and not the other way around either. In audio education there are different schools of thought when it comes to subgroups. Some say that subgrouping should be kept to a minimum while some say that subgrouping is a tool that should be used freely. The

(16)

The results of the listening experiment are most likely influenced by the way the mixes were created. If there had been less emphasis put on control in favour of ecological validity when designing the instructions, the results would have been different. One example that could change the results is to let the mix engineers use their whole toolset, i.e. the DAW that they usually work in, the plugins they usually use and other mix techniques. Since every engineers mixes were compared to their own in addition to the anchor, this would

probably not make a huge difference in terms of control. One way to change the listening experiment is to use a scale for preference for all subjects to use instead of letting subjects use their own interpretation of preference. This would have allowed for using the data as is, instead of having to use min-max normalisation.

The qualitative data from the listening experiment is very spread out and

responses were often vague. This could have been different if the questions for motivation of the ratings were thought out more. The question for each mix was “What in this mix made you rate it the way you did?”. This could be

substituted with first having a pre-study to find out what perceptual attributes should be asked for in the main experiment. Then the subjects could just use a scale to rate each attribute separately.  

(17)

5. References

De Man, B., Mora-Mcginity, M., Fazekas, G., & Reiss, J. D. (2014, October). The open multitrack testbed. In Audio Engineering Society Convention 137. Audio Engineering Society.

De Man, B., Boerum, M., Leonard, B., King, R., Massenburg, G., & Reiss, J. D.

(2015, May). Perceptual evaluation of music mixing practices. In Audio Engineering Society Convention 138. Audio Engineering Society.

Ronan, D., De Man, B., Gunes, H., & Reiss, J. D. (2015a, October). The impact of subgrouping practices on the perception of multitrack music mixes. In Audio Engineering Society Convention 139. Audio Engineering Society.

Ronan, D., Gunes, H., Moffat, D., & Reiss, J. D. (2015b). Automatic

subgrouping of multitrack audio. In Proc. 18th International Conference on Digital Audio Effects (DAFx-15). DAFx-15

Ronan, D. M., Gunes, H., & Reiss, J. D. (2017, May). Analysis of the

subgrouping practices of professional mix engineers. In Audio Engineering Society Convention 142. Audio Engineering Society.

Ronan, D., Ma, Z., Namara, P. M., Gunes, H., & Reiss, J. D. (2018). Automatic minimisation of masking in multitrack audio using subgroups. arXiv preprint arXiv:1803.09960. 

(18)

6. Appendix

6.1. Instructions (A) for the mix engineers

Mixes for listening experiment

Hi and thank you for participating! Your assignment is to create two mixes for one song in the genre indie-rock. The song is called Sea of Leaves by Jokers, Jacks and Kings and is 33 tracks large.

You will be doing the mixes “in the box” in your DAW. When it comes to plugins, you will be limited to using only one EQ and one compressor that I have selected and are free.

These plugins can be used as much as you want. No other eﬀects such as reverb or distortion may be used. Plugins used for optimal listening or measurement that sit on the master bus (such as sonarworks or a LUFS meter) may be used but must be bypassed at bounce. Work in the DAW you are most comfortable with and in the control room you normally work in.

Create projects with 24-bit as bit depth and 44.1 kHz as sample rate.

Set your DAW’s pan-law to -3 dB.

Plugins

Here is a link to download Melda Production MCompressor and MEqualizer:

https://www.meldaproduction.com/downloads

For this box in the installation, you only need to select MCompressor and MEqualizer under “Free Eﬀects”.

(19)

Subgrouping

Subgrouping is a mixing technique that involves sending multiple tracks to a bus before it is sent to the master bus. A subgroup can also be sent to another subgroup and that is called hierarchical subgrouping.

Mix A

In this mix, you may not use subgroups or buses at all.

Mix B

In this mix you may use subgroups or buses.

You will start with Mix A

Questionnaire

When you are done with the mixes, I want you to answer a short questionnaire here:

https://forms.gle/FUysZfRhPAxM32vE8

Delivery

Your mixes should be bounced as a .wav with 24-bit and 44.1 kHz and without dither.

Name your bounced files “Firstname_Lastname_Mix_A” and

“Firstname_Lastname_Mix_B”. Send them to me at the email address gb12128@gmail.com. I have set a deadline for March 12, 2021.

(20)

6.2. Instructions (B) for the mix engineers

Mixes for listening experiment

Hi and thank you for participating! Your assignment is to create two mixes for one song in the genre indie-rock. The song is called Sea of Leaves by Jokers, Jacks and Kings and is 33 tracks large.

You will be doing the mixes “in the box” in your DAW. When it comes to plugins, you will be limited to using only one EQ and one compressor that I have selected and are free.

These plugins can be used as much as you want. No other eﬀects such as reverb or distortion may be used. Plugins used for optimal listening or measurement that sit on the master bus (such as sonarworks or a LUFS meter) may be used but must be bypassed at bounce. Work in the DAW you are most comfortable with and in the control room you normally work in.

Create projects with 24-bit as bit depth and 44.1 kHz as sample rate.

Set your DAW’s pan-law to -3 dB.

Plugins

Here is a link to download Melda Production MCompressor and MEqualizer:

https://www.meldaproduction.com/downloads

For this box in the installation, you only need to select MCompressor and MEqualizer under “Free Eﬀects”.

(21)

Subgrouping

Subgrouping is a mixing technique that involves sending multiple tracks to a bus before it is sent to the master bus. A subgroup can also be sent to another subgroup and that is called hierarchical subgrouping.

Mix A

In this mix, you may not use subgroups or buses at all.

Mix B

In this mix you may use subgroups or buses.

You will start with Mix B

Questionnaire

When you are done with the mixes, I want you to answer a short questionnaire here:

https://forms.gle/FUysZfRhPAxM32vE8

Delivery

Your mixes should be bounced as a .wav with 24-bit and 44.1 kHz and without dither.

Name your bounced files “Firstname_Lastname_Mix_A” and

“Firstname_Lastname_Mix_B”. Send them to me at the email address gb12128@gmail.com. I have set a deadline for March 12, 2021.

(22)

6.3. Instructions for the listening experiment

Listening experiment instructions

Hello and thank you for participating!

You will listen to a number of diﬀerent mixes of the same song.

You will rate each mix according to your own preference. You will also write a short motivation for why you rated the mixes the way you did in this sheet.

At the top left corner of the software it says "Deltest_X". Write that number next to the respective "Subtest" in this booklet.

This is how the software works:

- The Play button plays sound.

- Pause button pauses sound.

- Position slider shows where in the sound you are listening.

- Start and stop sliders determine where the sound starts and ends.

- Press the loop button if you want the sounds to loop. Loops are determined by start and stop sliders.

- Select sounds by pressing the A, B and C buttons.

- Rate the sounds using the sliders above the A, B and C buttons.

- When you have rated all the sounds, press the next button to proceed to the next part of the test.

When you first listen, you can set your listening level using the green knob on the right, but you may not change the level during the rest of the test.

You can ask for help or cancel at any time. Your participation is voluntary and anonymous.

Thank you for your time!

(23)

Subtest ___

Your highest rated mix. What in the mix made you rate it the way you did?

________________________________________________________________________________

Your second highest rated mix. What in the mix made you rate it the way you did?

________________________________________________________________________________

Your lowest rated mix. What in the mix made you rate it the way you did?

________________________________________________________________________________

Subtest ___

________________________________________________________________________________

(24)

Subtest ___

________________________________________________________________________________

Subtest ___

________________________________________________________________________________

(25)

Subtest ___

________________________________________________________________________________