Expressivity in DJ performance

Categorising tempo variation in DJ performance

Chris Yeoward

Abstract

This project examines how tempo variation is employed by DJs as a tool for expression. Analysing sets by DJs across various genres shows three main patterns for tempo variation: static, dynamic and erratic. In comparing live and pre-recorded mixes, the tendency for DJs to be more conservative with tempo variation when playing to crowds also becomes apparent. Finally, the tempo patterns for two artists and the dynamic mix type are distilled into tempo contours, to provide a basis on which to guide real-time automated mixing systems.

Introduction

For many years DJs have had to fight their corner in the ring of legitimacy among musicians. An often heard snipe is that they are merely pressing play, and should take no credit for the music created by others. However, a DJs role is composed of several elements which require mastery both in and out of the DJ booth, and can be grouped into four stages: 

• Discovery: Searching for and collecting music, 

• Selection: Choosing tracks appropriate for the crowd, mood and atmosphere

• Mixing: Transitioning between songs smoothly 

• Programming: Structuring a set such that it progresses coherently, whilst remaining interesting

The rise of intelligent music streaming platforms has seen increased focus on the tasks of automated discovery and song selection. With the streaming industry continuing to grow, services now use sophisticated techniques to make informed recommendations to their users, hoping to improve a listeners experience and keep them engaged. The current state of the art uses combinations of usage and content based algorithms to make predictions [1]. Usage based models were limited in that they were better at promoting already popular songs and artists [1], whereas many avid music listeners are keen to hear music from new and upcoming artists. Content based models address this, looking at features drawn from the audio and music itself to group songs based on similarity [3]. The same concept is applied to playlists, grouping songs based on genre or mood to maintain coherence for the listening duration, and indeed the problem of song recommendation has sometimes been reframed as a music playlist generation problem [4][5]. However, consistency must come hand in hand with variety, as in [6] subjective assessments of playlists determined that similarity is only one of the features that listeners use to rate the quality of a playlist. Varying the artist and familiarity is also important, as is avoiding too many songs that are too similar. In [7], Cunningham et al interviewed different playlist curators to characterise the creative act of compiling a playlist. They found that playlists were initially created with a theme in mind, and then songs are added if they are not at odds with the current songs genres or tempos, or if the lyrics match the theme.   

Only recently has progress been made in automating the mixing component. Spotify recently added the ability for automatic mixing to one of their electronic music playlists: Housewerk, though research in this area remains light in the literature. In [8], the authors created a convincing real-time, automatic mixing system for Drum and Bass music, making assumptions about the structures of the songs to increase its performance. Successive songs were chosen from a randomly generated pool, with the selected song chosen based on the quality of the transition from the current song. Like other real time song recommendation systems, it lacked longer term structure, instead favouring immediate suitability between songs. As such, the question still remains as to how to structure a playlist or DJ set over a longer time period, maintaining coherence whilst also providing an engaging, captivating experience. 

This project seeks to understand how particular DJs keep their sets interesting over a longer term, examining tempo as a key feature of DJ expression. I use a variety of sets from different artists, both those recorded from a live performance, and those prepared privately for a mix series or podcast. 

Methodology

Data

The sets were taken from 'hottest' mix category from the online mix database https://www.mixesdb.com/. 'Hot' mixes were calculated from the number of favourites and page visits they received. Several restraints were placed on which mixes to use. They had to be:

• Approximately 1 to 2 hours in length (much longer DJ sets could be constructed completely differently)

• Recorded from a direct line in, not from a phone or camera.

• Performed by a solo DJ only, no pairs or groups

• Annotated with a track-list, with most track times provided

Track lists were required so that the performances could be analysed in track time, the number of tracks that have elapsed, rather than the time that has elapsed in seconds: performance time.

The first mixes to meet these criteria were selected. In total, 20 were chosen, of which 9 were live sets, 10 were pre-mixed, and one was recorded directly from Spotify's Housewerk playlist. 5 DJs were selected for both Live and Pre-mixed recordings, based on the availability of their sets:

• Hunee (house/disco/world)

• Four Tet (house/disco/garage/techno/drum & bass)

• Objekt (techno/drum & bass)

• Helena Hauff (techno/electro)

• Palms Trax (house/disco)

The sets chosen were:

The sets mostly focus on music around the 120 - 130 BPM mark, primarily featuring house, techno and disco.

Analysis

For a number of the mixes, the times for several of the tracks were missing, and so first required manual annotation. Subsequently, the track lists were parsed into a plain floating point format giving the start time for each track in minutes. The Essentia MIR library was then used to determine the tempo at each track time. A small margin was added to each track time to account for partially completed transitions, so as to ensure that only one track was playing at the time the tempo was taken. Three frames of approximately 20s each were taken from the track starting time, and used in the tempo calculation. The tempo was then measured as the average of these 3 measurements. Measuring sets in track time, and thus at discrete rather than continuous intervals, enabled a tempo differential to be measured for each set. This tempo differential that provided insight as to the rate of change of tempo over time for each set, and showed how quickly an artist would build or drop the tempo.

For each set, the mean, median, interquartile range, total range and standard deviation of tempos was calculated. 

Beyond this initial analysis, I desired to have some measure that could inform an automated mixing system of how it should change over time. This required selecting the most expressive sets, and reducing them to a normalised, average tempo curve, to give the tempo deviation from mean tempo over time, as a proportion of set duration.  Template tempo curves were then also generated for the two artists with the most mixes, Hunee and Four Tet, to give an approximate blueprint as to what an average set of theirs would look like.

Part I: Tempo Analysis 

Results

The figure below shows the tempo and tempo differential for the control 

Live sets

The plots of tempo and tempo differential against track number for each of the live sets are shown in the following figures:

Pre-mixed

The plots of tempo and tempo differential against track number for each of the pre-mixed sets are shown in the following figures:

The summary statistics for the tempo, tempo difference, and absolute tempo difference each set are shown below. Each plot shows the total range, interquartile range, mean (shown as the white circle) and median (shown as the diamond).

Discussion

From the above plots, there are three distinct styles that appear. There those that are static, and do not deviate much around the mean tempo. Examples of these include Rodhad's Essential Mix, Helena Hauff and Palms Trax's Boiler Room performances. There are those that are dynamic, that are gradually changing the tempo over time, as in the case of Hunee's mixes, Four Tet at Dekmantel, and Palms Trax's Essential Mix. Finally, there are those that are erratic, which are characterised by significant tempo changes over a short period of time, between one track and the next, often implying a dramatic change of genre. This behaviour can be seen in Objekt's mixes, Four Tet's essential mix, and Donato Dozzy's Alternate Cuts mix. 

The static mixes require no modelling as they simply never change in tempo. The dynamic mix type can be modelled by averaging over the tempo curves to determine what a generic mix would look like. This informed the second part of this study.

The erratic mixes, on the other hand, display discrete jumps between genres, and would be perhaps better modelled by a stochastic process. A model for these types of mixes could utilise a probability function that informs whether the mix will jump genre, as a function of time, current tempo or genre, and DJ. This would then aid in modelling Four Tet's mixes, which appear to be combination of dynamic and erratic, as is the case of his Lost Village mix. 

While many of the mixes display greater tempo variation than the control mix, not all of them do. This naturally leads to the question: why does a DJ employ one style over another? What causes Four Tet or Objekt to change tempo quite as much as they do, when other DJs such as Rodhad stay almost completely static? One explanation could be genre. Rodhad, Helena Hauff and Efdemin all play mostly techno, and all of particular types that hover around a similar BPM. There therefore is not much need to vary the tempo drastically, unlike those that weave between lower and higher tempo genres. A similar explanation could be for the purpose of the mix, and what crowd the set is designed for. Is it a warm up set, or is at the peak time of the night? Earlier sets could be expected to vary more than peak time, and this could also be linked with genre, as techno is more often played during peak hours. 

Live Vs Pre-Mixed

The link between mix purpose and tempo variation is displayed in the differences between pre-recorded mixes and live mixes. Without an audience, a DJ may be more inclined to stray from their usual BPM range, less obliged to keep dancers moving in a consistent manner. Four Tet exhibits this the most clearly. Whilst his live sets both follow an upward trajectory, with a break into another genre halfway through his Lost Village mix, his Essential Mix varies hugely. The Essential Mix has a mean absolute tempo difference of 10.7, in comparison to 2.53 for Dekmantel and 6.193 for Lost Village. Palms Trax also shows more variety in his pre-recorded Essential Mix. There's a clear upward trend, with a mean absolute tempo difference of 1.5 in comparison to 0.32 for his live mix.

However, Hunee shows greater variation in one of his live sets. His performance at Boiler Room has a mean of 1.32, at Brilliant Corners a mean absolute tempo difference of 5.7, and his Essential Mix 2.88. His performance at Brilliant Corners may have been an anomaly, or perhaps his Essential Mix was merely a typical performance of his, and is indicative of an average set.  

For the others, Helena Hauff showed a mean absolute tempo difference of 0.49 in her performance at Boiler Room, in comparison to her Crack Magazine mix which showed 1.56. Objekt's live mix was 8.82 and his pre-recorded 7.28. In general, the artists displayed greater variation in their live sets. 

Part II: Tempo Contour Generalisation

To characterise the DJ performances, and find an average tempo contour between them, each set was normalised and scaled to be a proportion of the deviation from the mean tempo along the Y axis, and as a proportion of mix duration along X. This was necessary due to the varied lengths of the mixes. An alternative approach could have been to measure the tempo in performance time rather than track time, and segment the mixes into hour long lengths. However, fitting them each to the same scale was preferred as a DJ set cropped to one hour will not be the same as a DJ planning for a one hour performance. As has been observed, DJs will often start and end at slower tempos, irrespective of the mix length. 

Four Tet and Hunee were chosen as artists to generate the templates for, as they were the most commonly featured. Each of the dynamic mixes were selected to generate a contou for what an average dynamic mix would look like. The selected mixes were:

• Palms Trax Essential Mix

• Hunee Essential Mix

• Leon Vynehall Resident Advisor Mix

• Four Tet Dekmantel

• Hunee Boiler Room

• Jeremy Underground Essential Mix

Results

Generalised Dynamic Mix

Hunee

Discussion

This section presented an averaged model for two artists, Hunee and Four Tet, and the dynamic mix type. 

The generic mix begins with a noticeable dip, though it is likely that this is caused by mid-tempo intro the prepended to all Essential Mixes. To account for this, future measurements would remove the intro. 

The curve rises gently, slowly plateauing to flat at around the 80% mark, before dropping in tempo for the finale, an often employed technique.

While the general trend for Four Tet's curve matches the upward trajectory seen in many of the mixes, the erratic tempo changes are likely to not be entirely representative of his behaviour. In averaging the values, the tempo's of these erratic changes have been brought closer to the mean, and therefore do not match those of the genres to which he deviates to, such as jungle or RnB. In these instances the model falls short, and would benefit from being combined with a probabilistic model to accommodate the genre switching. 

Conclusion

This assignment examined the tempo variation across various DJ sets, and presented three core types of tempo modulation: static, dynamic and erratic. A simplified technique for generalising the dynamic DJ set was employed, though there remains work to model the erratic song choices made by DJs. Studying an artist's corpus of mixes more closely may help train a probabilistic model to inform on the types of song choices a DJ may make. Another unknown is how much DJs are willing to distort the tempo of a piece of music to fit in with their set. If aiming for a particular tempo, how much is too much when it comes to tempo shifting? In addition, whilst several DJs displayed greater tempo variation in their pre-recorded mixes, a greater sample set is required to confirm this trend. 

Equally, there are other musical features that a DJ could employ as a means of expression. Timbral and rhythmic characteristics all play in to a DJs song choice, so analysis is required to determine the different ways in which they are varied. 

Finally, there is the mixing element of DJing itself. Whilst automatic mixing software has been developed, no work has been done to examine exactly how different DJs transition between songs, and how the transitions play in to the features of the songs themselves.

References

[1] Music Recommendation Based on Acoustic Features and User Access Patterns

[2] A Survey of Music Recommendation Systems and Future Perspectives

[3] A Content-Based Music Recommender System

[4] Understanding Music Playlists

[5] Contextualize Your Listening - The Playlist as Recommendation Engine

[6] HOW SIMILAR IS TOO SIMILAR?- EXPLORING USERS’ PERCEPTIONS OF SIMILARITY IN PLAYLIST EVALUATION

[7] Supporting the Creation of Playlists and Mixes

[8] From Raw Audio to a Seamless Mix - Creating an Automated DJ system for Drum and Bass