Deezer’s new AI software could change music recommendations

Reading Time: 4 minutes

Written by:

Music streaming service Deezer may not be the most popular platform in its field but has just made one of the most interesting developments the music industry has ever seen. The French platform has developed an AI system that can easily detect a song’s mood, an advancement that could change the way music recommendations are made by streaming services.

Founded in 2007 in Paris, Deezer remains one of the world’s most popular streaming services, having attracted a whopping 14 million active users each month as well as 6 million paying subscribers. The platform itself has over 53 million licensed tracks and 30,000 radio channels. With such a vast library it is unsurprising that the platform remains extremely popular. This latest move will undoubtedly see the platform experience even more growth, as it will enable Deezer to add even more accuracy to its recommendations.

Deezer researchers have now developed an AI system that is able to relate songs with moods and intensities. In order to discover a song’s mood, both the audio signal and the lyrics of the song were taken into consideration. Audio signals were fed into a neural network with models to reconstruct the context of the lyrics. Metadata from the Million Song Dataset was then used to help the AI learn how to determine the track’s mood, the metadata contained information on over 1 million diverse songs. Deezer researchers used Last.fm’s dataset, assigning identifiers to tracks using over 500,000 tags. On Last.fm, many of the tags are mood-relevant and over 14,000 words were rated based on their positivity and negativity. The words were also rated on how calm or energetic they were overall. Determining the positivity and energetic level of the tags allowed researchers to ascertain their general mood.

Rather than containing the songs themselves, the Million Song Dataset just contains the metadata, meaning that the researchers then needed to pair up their own catalogue with the data. They used specific details like song and album titles and artist names to link the songs with the metadata. The result was that 60% of the dataset was able to be used to train the brand new AI, this accounts for 18,644 tracks. The result of the research was that the AI was then much more able to detect the emotional mood of songs by analysing how calm or energetic and how positive or negative they were. The researchers stated:

“It seems that this gain of performance is the result of the capacity of our model to unveil and use mid-level correlations between audio and lyrics, particularly when it comes to predicting valence.”

The researchers also stated that in order to let the AI go further, it would be helpful to create a database with synchronized lyrics and audio. The other limitation is that not everybody feels the same way about songs, some listeners may find a song particularly sad while others may simply find it relaxing. A lot of determining a song’s mood relies on the listener themselves and the connection they make with the music. This makes the results subjective on a certain level. Despite the limitations, the new AI system is a huge leap forwards in the way that songs are understood and recommended. By understanding the positivity or negativity of a song, streaming platforms will be better able to differentiate between happy and sad songs, recommending them appropriately depending on a user’s listening history. This ability has unlimited positive effects for listeners and will make streaming platforms even more intelligent when selecting recommendations.

From an artists perspective, this is also highly beneficial. The more advanced streaming platforms become, the more people will be inclined to use their services. For this reason, artists could be set to receive even more streaming royalties through increased streams. Improved recommendation software is always positive for artists, as accurate recommendations mean that users are much more likely to find artists suited to their taste and actually listen and re-listen to tracks they are recommended. This gives artists a much higher probability of being linked to people that genuinely have a high likelihood of becoming a fan.

If your music is available on a streaming service, your tracks will be being recommended all the time, probably even more often than you realise. But, the truth is that many of these recommendations will be inaccurate and many of the users who are recommended your music may not actually enjoy it. This is the case for all artists and can be demoralising when there are plenty of people out there who would actually enjoy your style. For this reason, advancements in music recommendation are something that should prove highly beneficial for all artists. Believe it or not, the quality of recommendations on streaming platforms could have a huge effect on the number of streams generated by your tracks.

Ultimately, more effective music recommendations on music streaming platforms can provide endless benefits for your career. From linking you with people who are more likely to develop into big fans to generating more exposure for your releases, the possibilities really are endless. As technology continues to improve, the way we understand music is constantly developing and changing for the better. The more that we can understand music itself and the science behind it, the more effective we can be at recommending music to people who will actually enjoy it. This advancement is only the first step of many and highlights the way the future of music streaming is likely to look.