New Spotify Normalization Revelations

Hello Fellow Audio Nerds,

After our last town hall (over 2 months ago) I had several people reach out to share their experiences, expertise, and even give me (under NDA) a behind the scenes look at what they are developing in terms of problem solving solutions for the great normalization problem.

While I cannot share much of what was given to me in confidence, I can share a couple of extremely helpful points that can serve as puzzle pieces to complete our very incomplete picture when it comes to the mysteries concerning normalization.

As some of you might remember, much of what spawned the idea of coming together as a community in the first place was an experience I had with an important major label project. On release day, the artist swore his album was “a dB quieter” than everyone else’s music. I rushed to capture and test the streaming level of the album, and found that the album as a whole was testing at an integrated LUFS of -15.5. The artist was right!

That was when I introduced the idea of crowd sourcing data to start to try to crack the code on why some albums are louder than others.

Through the Audio Accountability community I was introduced to a new friend at a well known software company who offered some help. After explaining to him the situation, he confirmed on his end that the album was indeed streaming at an integrated LUFS of -15.5 and was puzzled as well as to why that would be. In a user created playlist, any song from the album would play at the expected -14.0 Integrated LUFS. However, when playing the album as a whole the volume would drop.

He reached out to a contact at yet another software company for help - and that is where we found our answer.

It turns out, Spotify does not calculate the integrated loudness of an album as a whole. It takes the loudest song, and reduces the whole album so that the loudest song sits at -14.0. This is a very important revelation.

For instance, if you have an album that more or less sits around -10 LUFS, and you have one song that for whatever reason (ie. dense song that has no soft sections) is sitting hotter, say around -8 LUFS Spotify will take the whole album down -6dB to make the -8 LUFS song -14 LUFS. The rest of the album that was more or less sitting at -10 LUFS pre-normalization will now be playing back at -16 LUFS.

If you mastered an album as a whole and the album normalizes low - check to see if any particular track is much louder that the rest. If it makes sense musically, turn the loudest song down a touch and you’ll find that the album as a whole will normalize louder than before.

Say for instance, we take the previous example where one song was -8 LUFS and the rest were more or less -10 LUFS. The -8 LUFS song normalized to -14 LUFS and the rest to roughly -16 LUFS. If we were to take the -8 LUFS song and bring it down 1dB to -9 LUFS, it would result in a 1dB increase overall in the normalized playback volume of the album as a whole.

Until next time,

Audio Accountability Town Hall 1

Spotify Volume Normalization (Known State and Issues)

Tuesday, January 17, 2024 - 5p PST

Participants (Names as appeared on Zoom)

• Dale Becker

• Pascal Pahl

• Rob Richards

• Matthew Teden

• Nick Kauffman

• Matt Huber

• Jeff Jackson

• Taylor Crommie (Crommatic)

• Mike Cervantes

• Edsel Holden

• Aram Piligian

• Jeremiah Douglas

• Eric Bard

• Connor Carroll

• Aidan Thillman

• Phil Feinman

• Noah Cronin

• David Turk

• Jason

• Josh Bowman

Post-Meeting Contributors/Commentators

• Tiago Loja

• Sam Moses (Moses Mastering)

Introduction: What is Audio Accountability? 

Audio Accountability is a multifaceted collective of music professionals committed to mastery and innovation in the audio arts. We seek to analyze and demystify audio from the recording process through release and beyond.

Our first objective as a collective is to compile a comprehensive pool of knowledge relating to music streaming with an emphasis in sonic fidelity and normalization (loudness) processes. Cutting through the misinformation widely available on the internet (some of which is from the streaming services themselves) we can better understand what we are hearing, why we are hearing it and what can be done to reach for a more equitable and high fidelity future in the music marketplace.

Current Goals of Audio Accountability. 

• Community and Interaction (getting us out of our studio caves) is an end in itself

• Inspiring all engineers to take ownership over the quality control over their work once streaming. 

• Developing a centralized place where relevant and easily digestible information can be found about audio.

• Although Spotify streaming loudness is our current & first subject to investigate, it will hopefully not end there. Future community endeavors can include:

• Streaming Codec Quality

  • Analysis of different streaming service codecs

  • Sample Rate Considerations

  • Re-imagining clearer language (“True” Peak, “Lossless” Audio)

• Artificial Intelligence “AI” Mastering Discussions (contributed by Tiago Loja)

• ATMOS Mixing and Mastering

• Community deep dives into certain plug-ins (ie comparisons and feature set explorations of common limiter plugins, etc.) 

• Curation of reference playlists for community use. 

• Publishing of Opinion/Review Articles from our community on our website

First Subject To Investigate: Spotify Normalization 

Spotify's purported methodology:

• Tracks are adjusted to -14 LUFS based on the ITU 1770 standard. 

• Anything below -14 LUFS is turned up to -14 LUFS as long as the -1dB True Peak ceiling is not breached

• Anything above -14 LUFS is turned down to -14 LUFS. 

Current Spotify Mastering Recommendations: 

https://support.spotify.com/us/artists/article/loudness-normalization/

Loudness normalization means we don’t always play your track at the level it’s mastered.

• Target the loudness level of your master at -14dB integrated LUFS and keep it below -1dB TP (True Peak) max. This is best for lossy formats (Ogg/Vorbis and AAC) and makes sure no extra distortion’s introduced in the transcoding process.

• If your master’s louder than -14dB integrated LUFS, make sure it stays below -2dB TP (True Peak) to avoid extra distortion. This is because louder tracks are more susceptible to extra distortion in the transcoding process.

Dale's Personal Experience (Unedited Notes - My apologies for any Typos) 

• Basing normalization solely off of the ITU 1770 standard with integrated loudness does not lead to a unified loudness result. 

• Because the integrated loudness is one simple loudness average, it under compensates loudness on less-dense tracks (such as acoustic songs) and overcompensates the loudness on denser tracks. An acoustic track that has a -10 LUFS integrated loudness might have the same apparent loudness as a fully produced dance track that has a -7.5 LUFS integrated loudness. If both of these tracks are turned down to -14 LUFS Integrated, The acoustic track will end up with a 2.5dB apparent loudness advantage over the dance track. 

• Because the integrated loudness is one simple loudness average, does not take into account a song's internal dynamics, which can have a huge impact on the final normalized volume.

  • Example 1: A "loud" dance track that is the same driving dynamic for the entire duration of the song, where every section is relatively the same loudness, let's say -7 LUFS, will have a -7 LUFS average.

  • Example 2: A similar track, who has open sections in between the loud sections (ie. verses with just pads and vocals) might have loud sections of equal volume to Example 1 that are -7 LUFS, but soft sections that are -12 LUFS. The average LUFS will end up being -9.5 LUFS and the track with open space will end up having loud sections that are 2.5dB louder than the track that is "Loud" the whole time. 

• To game the Spotify system, people have been mixing anti dynamically and with tons of harshness to gain apparent loudness. 

• Tracks mixed super aggressively in terms of presence tend to have a higher apparent volume. 

• Tracks mixed with extremely loud intro/1st verses which decrease in the volume when a chorus hits have much more apparent volume at the outset - which captures a listeners attention better than a track mixed with natural dynamics.

• Measuring Spotify's streams in LUFS clearly shows that not everything is normalized to -14 LUFS. I have albums that I've measured that are -15.1 LUFS, 1.1dB lower than their purported standard. I have measured singles that are nearly -14.7 LUFS. 

• Spotify’s recommendations and widely used tool sets (ie. Loudness Penalty) while well intentioned and helpful, are only accurate to a degree and generally lead artists, producers, and engineers to the wrong conclusions.

Additional Considerations 

• The sound of density and compression are a part of the sound of modern music. In many cases a pop song mixed or mastered at -14 LUFS will not have the saturation or glue our ears have come to expect in a pop song. 

• NOT ALL SPOTIFY STREAMING VENUES ARE NORMALIZED - while normalization is on by default for listeners, it can be turned off. Furthermore, other platforms, such as streaming to Smart TV's, do not have normalization as an option at all. That means if you release music to Spotify at -14 LUFS, you will be nearly 5 to 7dB quieter that most music on the Billboard Hot 100 when streaming to non-normalized scenarios.

• I was speaking with Ruairi, a mastering engineer friend of mine, who also commented that a well produced arrangement does come into play here. Many times people wonder why their music doesn't sound and feel as loud as the latest Cirkut produced single and that's simply because producers like Cirkut have mastered the art of imbuing energy into their productions. Things are balanced in a way that allow the song to speak with max amounts of energy, intensity, and transient impact. . 

Our First Community Experiment

Normalized Spotify vs. the WAV masters that were provided

• Introductory Notes:

  • Please email recommendations/report typos/etc. to dale@audioaccountability.com before Jan 31, 2024.

  • a detailed, written set of instructions will be provided after I complete the integration of community input.. 

  • Actual project names will be provided to Audio Accountability, but before data is published on community submitted data, project and artist names will be swapped out with a pseudonym to maintain anonymity. In the case that issues are discovered, we don’t want to be a cause for alarm to artists, production teams, etc.

  • Participating engineers will also be given the option of anonymity.

  • Analysis Tools TBD - We are currently in talks with a company that wants to help. Izotope RX is the toolset currently employed.

• The main purpose of the experiment is to gather, organize and analyze data related to real world Spotify streaming normalization performance by a two part process:

  • Analysis of the WAV Masters provided to the DSP

    • Sample Rate & Bit Rate

    • Single, EP or LP Running Time (h:m:s:ms)

      • For EP’s and LP’s please export whole project as one long WAV file with the top of the first song and tail of the last song as the WAV beginning and end of the export for analysis.

        • Naming Convention: BitRateSampleRate_SourceWAV_ArtistName_ProjectTitle_ContributingEngineerName

        • Example: 24b441k_SpotifyNormalized_TheKinks_WhyInTheWorldLP_DavidMcNasty

    • Integrated Loudness of the project - Single, EP, or LP

    • Ceiling True Peak and Sample Peak

    • Final Limiter Data:

      • Which plug-in limiter was used

      • True Peak or No? Oversampled or No?

      • Where applicable, which “Mode” (Transparent, Punchy, etc.)

      • Where applicable, what lookahead. (Can be omitted if confidentiality is preferred)

  • Playback and Analysis of the Normalized Spotify Stream

    • Source Machine (Spotify Playing) sound card set to 44.1k / 24b.

    • Note Computer OS (macOS or Windows Version) and Spotify Software Version (Spotify Menu>About Spotify)

    • Spotify Settings

      • Setttings Menu:

        • Streaming Quality: High

        • Auto Adjust Quality: Off

        • Normalize Volume: On

        • Volume Level: Normal

      • Main Window Bottom Right Hand Volume Slider - All the way up.

    • Connect Source Machine to Monitoring/Analysis Machine via AES/EBU or S/PDIF. Ensure sound card mixer or Computer OS is not manipulating output volume.

    • Monitoring/Analysis machine set to 44.1k and Digital Clock set to AES/EBU or S/PDIF input

    • Perform a small test stream to check sound quality/volume is as expected

    • Place a LUFS meter in your DAW (exact plugin TBD)

    • Reset the meter and play the project in its entirety.

    • Analysis:

      • Short, general observations of sound quality in relation to source WAVS.

      • Record Sample and True Peak

      • Record integrated loudness for project as a whole.

  • Fill out Audio Accountability Community Experiment 001 Form with findings.

    • All uploaded non-audio data/findings, with exception to project names, artist names and in some cases, contributing engineer names, will be made public

Information and data we gather, for now and the foreseeable future, will be publicly available for anyone to reference and use to their own ends.

Final Miscellaneous Links from Jan 17th Townhall (still being updated):

• Hollywood Sapphire Group: https://www.hollywoodsapphiregroup.com

• Mastering the Mix “Levels” https://www.masteringthemix.com/products/levels

• Sonible - True Balance https://www.sonible.com/truelevel/ (contributor: Crommatic)

• Apple Digital Masters / AU LAB https://www.apple.com/apple-music/apple-digital-masters/

Community Input:

Thursday, Jan 18, 2024.

Sam Moses (Moses Mastering) emailed with some very helpful input and brought up the legality of actually capturing Spotify streams. As professionals with the intent to research, rightfully entrusted with the actual source audio, with no piracy intentions, I had personally not worried too much about capturing streams. However, after careful thought, with the best intentions of the whole community in mind, we are shifting our methodology over to a monitoring and metering-only approach so no audio is captured during our experiments. The change will be reflected in the Town Hall notes as well as in our future publication of the experiment procedure.

He also brought up one thing which is a good thing to note. “I'm curious about the plan in place to future-proof the research findings. Given Spotify's tendency to change settings without user notification, how can we ensure the findings remain relevant over time?’

My answer is that we can’t. As with any technology, change is swift - but that does not negate the value of learning what we can now… and trying to keep up thereafter.