The final product

Investigating Loudness Normalisation in Spotify and Youtube

Lately I have been trying to bus compress my tracks to get them as loud as commercially released songs. This is quite frustrating as the amount of limiting required is bad for the sound. This got me thinking about the loudness war/loudness race and I had a look into loudness normalisation systems in Spotify and YouTube to see just how loud I really need to make my tracks... It's actually very interesting so I'll post what I found here.

The Loudness Race

Loudness is used specifically and precisely for the listener’s perception. Loudness is much more difficult to represent in a metering system. (…) Two piecing of music that measure the same on a flat level meter can have drastically different loudness (Katz, 2007, pg. 66).

According to Katz (2007, pg. 168), when two identical sounds are played back at slightly different levels, the louder version tends to sound ‘better’, although this is only a short term phenomenon. This has lead engineers to compete for loudness to attract attention in the jukebox/on the radio when the playback level is fixed, which is achieved through the use of compression and limiting on the master bus. Unfortunately, this can lead to fatiguing and eventually unpleasant sounding records, or in the extreme case, audible clipping, as in the famous example of Metallica’s “Death Magnetic” album (Michaels, 2008). The loudness race has also been present in television (with advertisements competing for audience attention) and radio (with stations competing for listeners to choose their station) (Robjohns, 2014).

Loudness Normalisation

The first attempt to implement some kind of regulation on loudness was the International Telecommunications Union’s (ITU) published standard, ITU-R BS. 1770, which in 2006 recommends a loudness metering algorithm to act as “electronic ears” to control the playback level of television advertisements. The algorithm uses a 400ms integration time, a two-stage filter to simulate the frequency response of the head and ears, and creates an average reading for the entire duration of the track (see figure 1). A gate is used to filter out quieter sections, so that they do not affect the reading (ITU, 2015). 

The ITU standard has been put to use in television, so that users do not need to reach for the volume control when changing channel or in between programs. Once the loudness of a program has been measured, a static level change can be implemented if it is too loud or too quiet, in order to reach a specific ‘target loudness level’, which for television is recommended to be -23 LUFS (loudness units relative to full scale) (Robjohns, 2014).

Katz (2007, pg. 172) predicts and end to the loudness race as consumer listening moves on from the compact disc to iTunes, whose “Sound Check” feature allows for the normalisation of the loudness  of audio tracks when in shuffle mode, removing any uncomfortable jumps in loudness when listening to music from different eras or genres. This processing is also applied to iTunes Radio. According to Robjohns (2014), iTunes normalises to a reading of around -16 LUFS using an algorithm that appears to be broadly similar to ITU-R BS 1770, although the details of Apple’s method are not publicly available. Robjohns suggests that this target loudness may be too low, as some devices restrict the maximum listening level to due loudness concerns. This means that users may not be able to compensate for the reduction in level. 

Other music streaming services have also started adopting loudness normalisation algorithms. For example, Spotify has the option to “Set the same volume level for all tracks”, which by default is switched on, however according to Shepherd (2015c), the reference level is too high; to keep a consistent level, Spotify adds extra limiting which can pump and crunch in an unpleasant way. YouTube has also implemented a volume matching system for music, which according to the findings of Shepherd (2015b) does not use ITU-R BS 1770 as the readings fluctuate between around -12 LUFS and -14 LUFS.

Figure 1. An outline of the ITU-R BS 1770 algorithm.

Procedure

To measure how YouTube and Spotify are normalising loudness, I used a program called SoundFlower to route the output of my MacBook in to Logic X and out again, where I could load Isotope’s “Insight” plugin for metering, which helpfully has a preset for measuring loudness using the ITU standard. Also shown are true peak values, which will show how much each track has been turned down.

Figure 2. Screen grab of the measurement process

SPOTIFY 

Track	Loudness	Peak Level	Notes
Closer by The Chainsmokers	-12 LUFS	-2 dBFS	Electronic/Pop (2016)
Didn’t Mind by Kent Jones	-10 LUFS	0 dBFS	Electronic/Pop (2016)
Final Song by Mø	-12 LUFS	0 dBFS	Electronic/Pop (2016)
Sunshine by Radio Edit	-10 LUFS	-1.1 dBFS	Electronic/Pop (2016)
That Was Just Your Life by Metallica	-12 LUFS	-6 dBFS	From “Death Magnetic”
Claire de Lune by Debussy	-17 LUFS	0 dBFS	Classical piano
Spotify advert	-6 LUFS	+1 dBFS
Don’t Let Me Down - Zomboy Remix	-12 LUFS	-3 dBFS	Dubstep

Spotify’s system appears to be working, keeping everything between -10 and -12 LUFS bar a few exceptions. The first five tracks were taken from a “New Music” playlist; these tracks were released this year. It is interesting to see that two of these tracks peaked at 0 dBFS, implying that Spotify did not turn them down at all; they were mastered to hit this target loudness. “Closer” and “Sunshine” were only turned down by 2 dBs and 1.1 dBs respectively, meaning they were not significantly louder. The Metallica track was taken from the notorious “Death Magnetic” album which whose over-compression brought the issues of the Loudness Race in to the public awareness. This track needed to be reduced by 6 dB in order to hit the loudness target.

To really test out the system, I also played a piece of quiet piano music (Claire de Lune by Debussy) which gave a reading of -17 dBFS. This falls below the desired loudness, but in practice it didn’t sound out of place; you kind of expect piano music to be a bit quieter. This seems to go against the claims of Shepherd (2015c) who claims that Spotify applied its own limiting to dynamic music. Either this is not the case, or certain types of music are exempt. It could be that classical music has its own loudness target that is lower than for pop music genres, due to the lack of bus compression used in classical music.

What let the system down was the exemption of adverts from the normalisation; from Debussy to Spotify’s advertisement for an electronic music playlist was a jump of about 11 LUFS which was  very unpleasant. 

YOUTUBE 

Track	Loudness	Peak Level	Notes
This Is What You Came For by Calvin Harris	-14 LUFS	-4 dBFS	Electronic/Pop (2016)
Closer by The Chainsmokers	-15 LUFS	-4 dBFS	Electronic/Pop (2016)
YouTube advertisement for Sky TV	-12 LUFS	-4 dBFS	Electronic/Pop (2016)
Heathens by twenty one pilots	-13.5 LUFS	-3 dBFS	Indie/Pop (2016)
The Day That Never Comes by Metallica	-14 LUFS	-4.5 dBFS	From “Death Magnetic”
YouTube advert	-13 LUFS	-4 dBFS
Rachmaninov - Piano Concerto No. 2	-17 LUFS	0 dBFS	Classical piano with Orchestra

YouTube appears to aim for a lower loudness level than Spotify. YouTube music seems to fall in the range between -13 and -15 LUFS, which agrees with the findings of Shepherd (2015b) who quotes between -12 and -14 LUFS. The advantage of choosing a lower target is that there is now a smaller difference when you switch from pop music to uncompressed classical music. YouTube advertisements are louder, at around -13 to -14 LUFS, but the difference is not as much as Spotify. This gives a much better listening experience. 

Conclusion

The findings show that Spotify’s system aims for a loudness of between -10 and -12 LUFS. Limiting a track any further than this is would be pointless as it would just be turned down. The Metallica track had very audible clipping when played side-by-side with other tracks at a similar loudness. For release on YouTube, it would be better to aim for a level of between -13 and -15 LUFS to maximise the dynamic range in the track. For release on both mediums, I feel it would be best to aim for Spotify’s playback level, otherwise the track may sound quiet compared with other tracks, or subject to Spotify’s automatic limiter, if there is such a thing.

It is hoped that these systems will reduce the commercial pressure on mastering engineers to over-compress music at the expense of sound quality. In radio, there is still no published standard yet (Robjohns, 2014) and this is an important step in finally removing all demand for hyper-compression in mastering.

References

ITU, (2015). BS.1770: Algorithms to measure audio programme loudness and true-peak audio level. [online] Itu.int. Available at: https://www.itu.int/rec/R-REC-BS.1770/en [Accessed 17 July 2016].

Katz, R. (2007). Mastering audio. Amsterdam: Elsevier/Focal Press.

Michaels, S. (2008). Metallica album sounds better on Guitar Hero videogame. [online] the Guardian. Available at: https://www.theguardian.com/music/2008/sep/17/metallica.guitar.hero.loudness.war [Accessed 17 Jul. 2016].

Robjohns, H. (2014). The End Of The Loudness War?. [online] Soundonsound.com. Available at: http://www.soundonsound.com/techniques/end-loudness-war [Accessed 17 July 2016].

Shephard, I. (2015). YouTube loudness normalisation - The Good, The Questions and The Problem - Production Advice. [online] Production Advice. Available at: http://productionadvice.co.uk/youtube-loudness-normalisation-details/ [Accessed 17 July 2016].

Shephard, I. (2015b). YouTube just put the final nail in the Loudness War's coffin - Production Advice. [online] Production Advice. Available at: http://productionadvice.co.uk/youtube-loudness/ [Accessed 17 July 2016].

Shephard, I. (2015c). Why Spotify's 'set the same volume level for all tracks' option is back, and why it matters - Production Advice. [online] Production Advice. Available at: http://productionadvice.co.uk/spotify-same-volume-setting/ [Accessed 17 July 2016].