They hadn’t sung a single line. Their vocals were generated by artificial intelligence (AI). The machine has moved on from synthesising sounds to generating singing voices. AI is on song everywhere.
Anshuman Sharma and Aditya Kalway, two young producers assisting the music composer duo SalimSulaiman, recreated “Haule haule ho jayega”—the hit song in Shah Rukh Khan’s Rab Ne Bana Di Jodi—in the voice of Mohammed Rafi. “We came across some AI vocal models that could produce the voices of Indian singers like Kishore Kumar, Rafi and Sonu Nigam,” recalls Sharma. In September 2023, Kalway sang “Haule haule” in Rafi’s signature style, put it through a vocal filter and musically arranged it like a 1970s LaxmikantPyarelal romantic ditty. Its Instagram Reel got 2.6 million views and cheers from the music industry, including Vishal Dadlani, Shaan and Sonu Nigam.
In January this year, composer AR Rahman strode into the AI ring. He used AI to generate the voices of Shahul Hameed and Bamba Bakya — two singers with whom he had collaborated but died, prematurely, in their 40s—for the song “Thimiri yezhuda” in the Rajinikanth film Lal Salaam.
AI is the new sound of music. A harbinger of change as well as confusion, it is shaking up the industry. While the technology’s potential is immense, artists are admittedly wary of their voices being cloned with the click of a few tools.“2023 is when the AI-hype phase in music really took off,” says Valerio Velardo, a consultant in music and AI based in Andalusia, Spain. He founded Melodrive, an AI system that composes music in real time, in 2016, followed by the Sound of AI, one of the largest online AI music ecosystems. “We shifted from a symbolic generation to an audio-based generation,” he adds, referring to the shift from the use of symbols to create music—which a synthesiser or sampler does —to the use of audio files to train AI models to generate music. “This was a leap,” he says.Indeed. It has come a long way from 1957, when the American composer and chemist Lejaren Hiller and the mathematician Leonard Isaacson programmed the Illiac computer at the University of Illinois at Urbana– Champaign to compose a string quartet. The Illiac Suite is widely regarded as the first score to be composed by a computer.
The use of artificial intelligence in music exploded with the entry of tech giants and powerful AI-first firms. One of the pioneers was OpenAI’s Jukebox. Released in 2020, it generated raw audio that approximated the inputs you gave like a certain musician or a particular genre.
Google came up with MusicLM, which generates hi-fi music from text prompts, in May 2023. Meta followed with AudioCraft in August. Soon, Stability AI, which created the popular text-to-image model Stable Diffusion, swept the music industry with its Stable Audio. A text prompt like “a low-key romantic ballad in a Scandinavian setting, with harmonies across three octaves in the style of Charlie Puth” would generate exactly that, in a voice and style uncannily similar to the American singersongwriter. “This is the equivalent of Midjourney or Dall-E, but for music,” says Velardo.
In addition to Big Tech models, there is a proliferation of free, easy-to-access AI tools for music, especially voice-cloning applications. Sharma and Kalway were playing around with such tools when they decided to bring alive the sound of Rafi.
WHO OWNS MY VOICE?
The easy replication of famous voices with the use of AI has raised a significant question—who holds the copyright to a voice? Its particular timbre and texture? That is sending ripples across the music industry, leading to legal tangles and conversations about copyright.
When Universal Music Group, the world’s leading music company, got into a licensing dispute with shortform video company TikTok Inc earlier this year, AI-generated music was a big part of the quarrel. Universal said TikTok was “flooded with AIgenerated recordings”. It pulled down ghostwriter977’s “Heart on my sleeve”. Before that, Universal had sued AI company Anthropic for distributing copyrighted lyrics with its AI model Claude 2.
Singers, too, are raising their voices against the use of AI. Earlier this month, 200 artists, including Stevie Wonder, Robbie Williams, Billie Eilish and Katy Perry, wrote an open letter, asking AI developers, tech companies, platforms and digital music services to stop the usage of AI in music which, they said, “diminishes the rights of human artists”.
The ripples have reached India too. In January, SP Kalyan Charan, son of the legendary singer, S P Balasubrahmanyam, sued the producers and music director of a Telugu film called Keedaa Cola for the “unauthorised” use of his late father’s voice with the help of AI. The filmmakers have denied this. Singer-musician Shankar Mahadevan is cautious. “Use of technology is always a fine balance between possibilities and restrictions. AI in music is an amazing advancement in technology,” says Mahadevan, who is part of the Grammy-winning fusion band Shakti. “But it should be used to enhance a basic composition and not create or replicate human creations. Otherwise, it could be like copying someone’s tune. Is that ethical?” he asks.
The surge of AI has many ramifications. It could, for one, eat into the chances of a new singer. Today, new artists, looking for a break, render scratch versions of a song before the final version is sung by an established artist. It is these scratches that bring new artists to the notice of music directors. With vocal cloning, all that a music label needs to do is choose from a list of vocally filtered renditions of established singers to decide whose voice will suit the song. The trial-and-error of scratch versions, often the first step to fame for a new musician, could get completely bypassed.
“I think AI is a good tool in the hands of somebody who understands the organic creative process and the electronic creative process,” said tabla maestro Zakir Hussain and founder of Shakti in a recent interview to ET, adding, “In the hands of someone who is totally uninformed, I don’t know what it will become.”
AI can bring tremendous opportunities to the world of music. A recent report by EY said, in terms of revenue growth and cost benefits, generative AI would have an impact worth `45,000 crore in the media and entertainment business in India over the next few years. “It could mean a 30-40% enhancement of music industry’s revenues,” says Ashish Pherwani, media and entertainment leader, EY India.
“Now, we can keep alive yesteryear’s artists in the minds of their fans for longer. The cost of recreating a fan base vanishes because the fan of Hemant Kumar or Rafi will listen to the songs you create. It’s a dream-come-true from a business point of view,” he adds.
LEGAL VOID
There is a yawning gap, though. There is currently no law that protects artists, their voices, vocal styles and compositions from being recreated by AI-generated tools, says Velardo. “It’s a legislative void,” he says. He points to two problems: “One is straightforward: I’m recreating Taylor Swift’s voice. That’s a breach of copyright. There’s another that is more subtle, which is training full generative music systems. AI companies train their models on material they don’t own and they don’t have permission from copyright holders to do the training.”
The absence of laws is a problem that Charan’s lawyer Kavitha Deenadayalan is facing in the lawsuit against the makers of Keedaa Cola. “There is a narrow difference between using AI-generated music for the sake of entertainment or for paying tributes to a late singer, and using it for commercial purposes.
When it is commercialised, somebody has to get licence and consent and also pay the artist the remuneration that is mutually agreed upon,” she says. “In the absence of a law, there is an ethical need to reach out to artists or their family for permission. That’s what Rahman did for ‘Thimiri yezhuda,’” she adds. Gautam KM, partner at Krishnamurthy & Co and an expert in AI law, says, “When you are generating music in someone’s voice through a platform, where was the platform trained? Was it, in this instance, trained on the deceased singer SPB’s existing songs? If that is the case, has the platform owner taken necessary rights from his estate or heirs to train the platform with SPB’s voice? If you are giving lyrics to a platform and asking it to create a sound recording in the voice of SPB, you could very well be infringing upon the personality rights or the moral rights of the original singer,” he adds.
Legal guardrails are coming up. The state of Tennessee in the US has enacted a new law called the Ensuring Likeness, Voice and Image Security (Elvis) Act of 2024, which seeks to impose liability on AI and tech companies for unauthorised use of a person’s voice or likeness. The law will come into effect on July 1.
Meanwhile, artists are putting up their own guardrails against AI. Canadian singer Grimes has created an AI software called Elf.Tech, which allows users to clone her voice but they will have to share 50% of royalties with her. “The moment you commercialise a song with Grimes’ AI voice, you are going to do a revenue split,” says Velardo. She has also tied up with New York-based music distribution company TuneCore to distribute the songs. In a recent interview on The Music Podcast, singer Arijit Singh spoke about copyrighting his voice. Mahadevan echoes that view. “Every artist has to copyright his or her voice. I certainly would,” he says.
