After almost a decade of focusing its operations on music-synchronization licensing, Rightsify Group LLC is pivoting operations towards an artificial intelligence-powered music generation tool called Hydra.
The Pasadena-based company said it’s working not only to support human-made music with this new tool, but to help advance and train AI models across the market to improve audio-generation abilities. After initially launching its generative AI service in December, Rightsify recently released a new, advanced version.
Since its founding in 2013, Rightsify’s operations previously focused on licensing music for television shows, advertisements, film and video game scores. Over the years, it expanded into licensing its catalog as background music for public spaces. Alex Bestall, Rightsify’s chief executive and founder, said that that 95% of Rightsify’s music library is instrumental, adding that its public space licensing is most popular with airports, hotels, airlines, spas and retail stores.
Trainable data is new model
Rightsify purchases rights to music from artists outright. In the process of collecting information on its catalog, Bestall said that Rightsify “accidentally” built a massive database library that tagged songs for details such as key tempo, chord progression, instrumentation, song structures, time signature and genre.
Bestall said a “big tech company” reached out to Rightsify early in the Covid-19 pandemic to inquire about using that database to train an AI model.
Rightsify originally rejected this offer, but after watching AI’s importance and capabilities grow in recent years Bestall realized what an asset the company had on its hands.
In the last year, licensing AI music datasets to other tech companies has become Rightsify’s predominant revenue stream. This year, the company plans to open source its datasets. Rightsify has never pursued or received venture capital funding, which Bestall said has allowed the company to be flexible and independent, and the company said it has become profitable on its own.
“After the launch of (OpenAI’s) ChatGPT in 2022, we saw that, obviously, this could be very disruptive,” Bestall said. “Instead of waiting on the sidelines or fighting it, we really embraced it. We’ve been licensing our music as datasets to tech companies for a little over a year now. While we had been preparing datasets, we had learned a lot about machine learning and how these models work. We started training our own, which led to Hydra I in December.”
Generation capacity for music
While generative AI’s ability to produce realistic text and images has advanced tremendously in the last few years, music generation has lagged in development. Music is arguably one of generative AI’s current weak spots due to the medium’s complexity and the lack of trainable, open-source data available to build models from. According to a recent report from Meta, the challenging nature of AI-created music comes from the huge variety of complex signals and the local and long-range patterns that music itself is built on.
“I assume the models (on the market) will get a lot better this year,” Bestall said. “What’s come out in the previous two years was based off of limited data… those datasets were basically just song name, genre, tempo and a description. But when you can actually map things down to the chord progressions, the bars and the instrumentation for each part of the song, (the models) can get a lot better.”
Hydra’s interface allows the user to enter a musical request and receive a generated instrumental sample between 10 seconds and two minutes in length. The more detailed the request, the better: Hydra will create a more relevant and complex musical sample if asked for “lo-fi hip-hop music with low bass at 85 beats per minute,” than it will if asked for “slow hip-hop music.” Hydra I was launched in December, and Hydra II launched last month.
In addition to a free model, which allows up to 10 generations, the platform offers a $39 per month “professional” subscription with 150 generations and a $99 per month “premium” subscription that includes 500 generations. The professional and premium subscriptions include unlimited remixes of generated content and, unlike the free model, allow the generated music to be used commercially.
The difference between the two lies in the editing and mixing capabilities provided to the user, and in the size of the AI model’s knowledge. Hydra II is trained on 1 million songs, while Hydra I is trained on only 60,000 songs. Bestall said that the smaller dataset means that Hydra I has less diversity of genres, instruments and compositions in its generated samples.
Hydra I allows users to generate and then download clips. With Hydra II, users can edit music by looping clips, changing the tempo, pitch or key, pulling out specific instrument components, mastering and more. Hydra has about 10,000 users combined between the first and second generation, which is nearly double from its user count in the first week after Hydra II launched.
Neither model includes any vocal-generation components, and only instrumental audio can be created. Bestall said that Hydra’s purpose is creating background music, instrumentals and backing tracks. As a result, the company doesn’t see AI-generated vocals as a necessary offering for Hydra to succeed.
“We still have a sizable amount of vocals (in our licensing catalog), but it’s the area with AI that there is the most potential for abuse, so we just want to steer clear from that entirely,” Bestall said. “We do license vocal datasets to other tech companies, but in terms of having it in our (AI) model we don’t see it as a great selling point for the customers we’re trying to reach.”
AI music’s purpose and place
Rightsify’s music-licensing business and its audio-generation platform both face competitors in the field, including generative AI tools such as Massachusetts-based Suno Inc., Meta’s MusicGen and OpenAI’s MuseNet. On the music-licensing front, another contender is Santa Monica-based Songtradr Inc. According to Songtradr president Bryan Biniak, AI-created music and human-created music fall into very different parts of the market. He said that the former serves a specific commercial purpose in the product and stock music market landscape.
“The key thing to remember is that AI can never replace the authentic human connection that artists bring through their work to their communities of fans,” Biniak said. “Whether it be through tours, meet and greets, social media, merch and more, part of what makes music so impactful is the emotional (response that) brands license music for: to connect with the audience that an artist has created. AI-created music can’t replace that.”
Joe Poindexter is an adjunct professor at the USC Thornton School of Music, as well as the chief communications officer and executive vice president of digital at Silver Lake-based Pulse Recordings. Poindexter said that, from an ethical stance as well as an audience-engagement point of view, it’s crucial to consider where AI-generated music may be used. An audience likely won’t mind much if AI-generated music is used as an “underscore,” meaning as background music in a public space or piece of social content. However, if AI-generated music is used in a film or television score, Poindexter said the audience engagement may lessen and human musicians’ financial prospects could be impacted.
Bestall said that he sees Rightsify’s services staying more in the business-to-business space, rather than being used for television shows or ads.
“I think we have to be confident that AI instrumental music or underscore music is a threat to composers who make a living off doing this,” Poindexter said. “I think we also have to be optimistic that the entertainment industry is going to protect human creation, because they just fought to protect it for themselves (during last year’s strikes).”
Biniak said that every new technology has the potential to change markets, and that AI is no exception.
He stressed that significant changes will occur as the volume of generative AI-created production and stock music increases in the market, which will translate in the commoditization of the category. That could result in reduced market demand and market value for human-made music.
Unlocking the ‘democratization of music’
Biniak and Poindexter both said that artists’ abilities and opportunities could also be expanded as new business models are built around the licensing of AI creation tools – such as Hydra.
Musicians that are just starting out or have limited resources can add AI to their toolbox of production options, on top of music creation software programs such as GarageBand or Pro Tools.
“What this will unlock is the democratization of music composition and production, creating opportunities for non-professional musicians to use generative AI to create and commercialize music in a way that hasn’t been available before,” Biniak said. “Given this blending, at the moment it’s impossible to say what long-term financial impact AI will have on human-made music, but … human-created music is integral to the music industry.”
Going forward, Bestall said that he anticipates that dataset licensing will remain the largest part of the company’s operations.
He added that human-created music will continue to have value for both the music industry and for AI development itself.
“We’re still making a lot of human music every single day, because you can’t feed AI models entirely AI music,” Bestall said.