Suno is the fifth most used generative artificial intelligence service in the world and continues to make technological breakthroughs, even as it faces a lawsuit from the recording industry for using copyrighted songs to train its AI model. The latest version, V4, is significantly more realistic and is already available to paid subscribers, before being rolled out to all users. «It's becoming a platform that generates songs that I want to listen to», says Mikey Shulman, co-founder of Suno, «and not just experiments to improve».
We are in a brand new recording studio (with guitars, basses and a top sound system) set up in the company offices, which are equally new. They occupy two floors, soon to be three, near the Harvard University campus in Cambridge, Massachusetts. “We were forced to improve the model to justify purchasing these top-of-the-line speakers,” Shulman jokes. In February, the company had a dozen employees, now it has more than 50, and more are on the way. “It's difficult to compete with OpenAI when it comes to getting the best researchers,” explains Shulman, mentioning the artificial intelligence giant behind ChatGPT. «But we have our own strength: we are the best when it comes to adapting AI models to the tastes of human beings».
If large language models can be judged according to objective parameters (for example, you can compare the scores of Claude and ChatGPT on college admissions tests), Suno engineers have people's tastes as a reference point . And so the model was improved by taking note of the preferences expressed by users who found themselves faced with different results starting from the same prompt. “Now, after a few months of work, we have a more precise idea of what human preferences are.”
The music generated by artificial intelligence from Suno or its closest competitor, which is Udio, does not have a full sound and in this it resembles that of old MP3s with low bitrate, something that is particularly evident in the vocals. We spent a couple of hours generating song after song. Result: V4 creations have never looked as good, with more realistic singers and instruments and a stronger stereo effect. Shulman says the model has also improved in songwriting. «Music is becoming more interesting. There are chord sequences that one wouldn't expect.”
For one of our attempts, which you can listen to below, we used a lyric I wrote on the spot paired with a prompt that asked for a song by organic country. The result is impressive, you can almost see the worn edge of the cowboy hat worn by the fake vocalist. The singing seems slightly filtered by AutoTune, probably due to the many electronically modified voices with which Suno was trained.
For those who hate artificial intelligence, a category of people that at least on paper includes almost the entire recording industry and its artists, many of whom have signed petitions against its use, the prospect of an AI even more effective in generating music is bad news. There are, however, exceptions. Timbaland, for example, told Rolling Stone who used Suno 10 hours a day to refine unfinished songs and who worked with the company as a creative consultant. According to Shulman, many artists, writers and producers happily use Suno. Among them there is at least one star who has signed a petition against AI.
Shulman hopes that an agreement can eventually be reached on using data to train AI and believes that artists should be more concerned about other models, ones that will reproduce their voices in the future even if they have not been specifically trained with the AI. their material. Suno doesn't allow you to do this, as it prohibits the use of real artist names in song generation instructions. “Sooner or later, without using Neil Young's music, someone will train a model capable of creating an exact copy of Neil Young.”
Now Suno no longer simply generates songs based on short typed lyrics like ChatGPT. It is possible to upload your own unfinished compositions, vocal-only tracks, loops and other audio fragments and transform them into songs, reaching a very advanced level of collaboration between humans and AI (there is even a function available in beta version that allows you to upload videos or photos to inspire the songs).
Rebecca Hu, project manager at Suno, says the ability to work on pre-existing audio is attracting the attention of young beatmakers. «Many of ours power users they are young producers. For them it represents the future. We're trying to move to a UI that responds to music and not text prompts, which are harder to interpret when it comes to generating sounds. Producers and authors who are in the studio working on their songs will make the most profitable use of it.” For now, the company is still focused on its original mission of enabling non-musicians to make music.
V4 also offers the possibility of using a new model, currently being refined, capable of generating more credible texts than the previous ChatGPT model. He's significantly better at writing rap lyrics, although in one of our tests he copied Drake's 2015 line, “running through the Six.”
Most of the company's employees aren't worried about the copyright infringement lawsuit looming over Suno, Shulman says, but “obviously it weighs, because nobody likes to be sued. Here, however, we are building the future of music. The lawsuit is an obstacle, but it will not stop us from building that future.” The next step is to involve labels and artists by making them partners, “because we cannot, nor do we want to, build that future alone”.
From Rolling Stone US.
