“The samples created by Google’s MusicLM is great in terms of diversity and fidelity. There have been other audio-generating software like Riffusion and others, but this one stands out. This can also be due to the database that Google has access to. The only concern that I have is the way the data has been sourced. It’s a grey area,” says Rekoil Chafe, one of the original lyricists and music producers of the Mumbai hip-hop scene. He is currently experimenting with Artificial Intelligence (AI) tools for his upcoming music projects. Google’s upcoming AI music tool MusicLM is also a project that Chafe has been closely looking at.
Before tech enthusiasts could move on from OpenAI’s ChatGPT, MusicLM became a topic of discussion among AI and music enthusiasts on Twitter.
Those ‘wow’ moments are happening daily.
Create a song without any musical know-how.
Have you ever hummed a melody & thought, “That would be a good song”?
Enter Google’s MusicLM AI.
Here are some examples 🧵. Starting with the hum prompt 👇[Turn the sound on each video] pic.twitter.com/OdYm5rbIIW
— Dan Fitzpatrick (@DanFitzTweets) February 3, 2023
This January, Google researchers released a paper introducing MusicLM, an AI model for generating “high-fidelity” music, which is basically high-quality music, primarily from text prompts. The results generated range from seconds of music to minutes-long pieces. Google has publicly released MusicCaps, “a dataset composed of 5.5k music-text pairs, with rich text descriptions provided by human experts”. These text captions describe the music and list out a few musical aspects describing the “genre, mood, tempo, singer voices, instrumentation, dissonances, rhythm etc”.
One of the interesting features, demonstrated in the samples is that the MusicLM can be “conditioned on both text and a melody”, in which whistled and hummed melodies can also be transformed to music pieces according to the style described in the text description. Chafe recommends listening to results from using ‘Bella Ciao’ as the melody prompt with text prompts that basically direct the instrumental sound for the same melody. The author’s personal favourite includes the audio generated using painting descriptions as prompts, particularly the music of the ‘Dance’ by Henri Matisse.
You can listen to the samples here.
Google claims that MusicLM outperforms existing systems in terms of audio quality and adherence to text descriptions. The researchers have evaluated the results by comparing it to Mubert, another AI platform for music creation and distribution, and Riffusion, which generates music from images of sound and is based on open-source AI model Stable Diffusion that generates images from texts.
STAY ON TOP OF TECH POLICY: Our daily newsletter with top stories from MediaNama and around the world, delivered to your inbox before 9 AM. Click here to sign up today!
Now, the problems:
An intervention by generative AI tools in any sector opens up risks of infringement, copyright claims and privacy threats too. Recently, Getty Images sued Stability AI, creators of Stable Diffusion, on the grounds of infringement on their copyright. It alleged that the AI company copied over 12 million images from their database without Getty Images’ permission or compensation “as part of its efforts to build a competing business”.
In 2022, a class action lawsuit against Microsoft, GitHub and OpenAI was filed by copyright lawyers who alleged that GitHub Copilot, an AI-based coding assistant, has been found to use “long sections of licensed code” without crediting the original coders. The case is said to have a huge impact on the use of artificial intelligence and its implications on copyright laws in various sectors.
Google too is currently cautious about potential legal challenges. In its MusicLM introduction paper raises concerns of “biases”, “cultural appropriation”, “potential misappropriation of the creative content to the use-case” and emphasises the need for more research. “We strongly emphasize the need for more future work in tackling these risks associated to music generation — we have no plans to release models at this point,” it adds.
Just as Chafe pointed out the concerns related to data that is used for training MusicLM, the copyright lawyers that the author spoke to for this article, highlight similar issues.
What do Copyright lawyers say?
1. Who is the author and owner?
The ownership of an AI-generated work is a key point of discussion when lawyers deal with copyright cases. Until the tool is used for a personal use there’s not much of an issue, but when it comes to distribution or using these works for wider access, then it gets trickier.
According to Gautam KM, a media and entertainment lawyer, it is necessary to understand that under the Indian Copyright Act 1957, there’s the author, who has created the work, and then there’s the Copyright owner. If a creator is not appointed by any person or group, they are the author and owner of their copyrighted works. In instances, where a creator is appointed by a company or a record label, in musical works’ case, or a studio, then the creator is considered the author and label or the commissioning entity is the owner. This is the distinction, which came about in the 2012 amendment and is fundamental to analysing copyright cases, informs Gautam. Section 2(d) and Section 17 of the Copyright Act define these categories.
In terms of AI, Gautam explains that with the interpretation of the Copyright Act, “for all practical purposes, the AI is the author” because the AI too will be creating compositions, lyrics or other musical works. But, the question is how do we enforce a law on a tool, because the law says a real person is the author. According to him, in such cases, the creator of the AI tool is the author and the company appointing the creator, in this case Google, is the owner.
Given that royalty related legal battles in the music industry affect more producers, composers, lyricists, distributors and other stakeholders, the lawyer states it will be interesting to how AI will disrupt this space. “There are a lot of grey areas under the Copyright Act itself which need to be addressed. One such is on the royalties itself. When a tool is creating the work and the author is not an actual person, how would we see royalties in that sense then? Currently, the way it stands, it works simply that the person who created the tool will get the royalties for the work,” he adds.
He observes that for AI-specific cases, it is clear that the distinction of author and owner needs to be assessed. “Where the challenge will lie is when it comes to contracts and the negotiations between the parties. But I don’t see any particular challenge as such when it comes to AI and the way we would interpret the Copyright Act currently,” he adds.
2. Proving the malafide intent:
Gaurav Bhalla, a media and entertainment lawyer from Ahlawat Associates, highlights that proving the malafide intent to copy a work will arise as one of the challenges in an AI-related copyright case, which can be dealt with under the Copyright Act in conventional cases. In terms of MusicLM, he explains, “Now let’s say a song which pre-exists in the real world but was not taught to this AI, right? But the AI knows the melody, the AI knows different beats and through a combination of those it is able to create a musical work which is similar to another song which wasn’t fed into the AI. So in such a scenario, of course, proving that there was malafide intent is difficult”. He opines that even in such cases, the copyright claim would sustain over the AI, but there are pre-enforcement challenges which will be encountered by the lawyers.
3. Where is the data sourced from?
“I think we need more visibility into what are the data sets that are being used. Datasets used for Machine Learning need to be licensed, but when it comes to MusicLM, where is Google getting a data set from? Are these data sets that have been licensed from musicians or artists or rights holders who have the necessary rights to pass that on to Google?” questions Sandhya Surendran, an entertainment, tech and media lawyer.
Referring to the Indian AI music platform Beatoven.ai, Surendran says the platform buys out samples, loops from artists and the exchange of creative works is properly commissioned with restrictions in place with respect to distribution and compositions. According to Surendran, the reference material used to create MusicLM still has some kind of human inputs for composition that’s being fed for Machine Learning. She states if Google uses Google Music for sourcing data, there would be a clear violation of the intention behind a distributor supplying music to Google Music for streaming purposes, as opposed to using the music for data learning purposes. “Unless that is the arrangement, then we are looking at a whole new set of problems,” she adds.
4. Will AI works be classified as derivatives?
And what if a music producer or artist wants to use MusicLM as a reference point for creating a derivative work? A derivative work is simply a work based on one or earlier works such as musical arrangement, sound recording among others. Surendran says using MusicLM as a reference point to create a derivative work, will also bring in the licensing aspect to understand whether Google can pass on the right to create derivative works too. This is will also affect artists will do sampling of a song, which is basically reusing a portion of a song in a new work with or without adding extra musical elements to it.
According to Surendran, this is an unaddressed issue. “We need to explore whether machine learning and the results from machine learning based on certain data set will also qualify as derivative. I would think it is, but then I feel that that kind of makes it virtually impossible to commercially exploit something like MusicLM, because then it’s a copyright infringement,” she adds.
5. No immediate worries
Though generative AI tools are quickly making their way in various fields and the Indian government is already tracking the developments in this space, lawyers say its tremors have not been felt in the Indian music industry yet. However, it is evolving at a fast pace and lawyers expect legal challenges to arise in the coming four to five years.
“I don’t think we’ve reached that stage yet. Mainly because every platform, whoever is dabbling in AI music production, is very clear that this is not for commercial use. This is for very specific personal use, for select set of users, or for maybe to sing for a podcast or a theme for a video. You can’t distribute the song as is,” says Surendran adding that it’s only a matter of time though when things will start changing.
What does it mean for artists?
Chafe is particularly intrigued by the capability of MusicLM to create instrument melodies from text and humming prompts too. Given the tool’s ability to understand instruments fairly good enough, he believes this can be used in future for separating orchestral sounds from a song, which is currently done using tools such as Audio Splitter. according to him, such sounds produced by an AI tool can also become a part of the YouTube Music library, ultimately affecting the royalties of artists who have already submitted their works for the library.
“This will create a saturation point and will cause a dent in the earnings of these artists, who are already struggling against big labels even in the Spotify streaming space,” says Chafe.
While apprehensions about generative AI tools taking over people’s jobs cannot be discredited, Chafe believes it will also enhance accessibility to new age tools for creating art to a lot of artists from the marginalised communities in India. Amid lack of access to resources, industry networks and costly sophisticated tools, softwares like these enable artists to enter the creative space by building upon their fundamental knowledge of how internet works.
“That kind of accessibility is open. Again, along with that accessibility, there needs to be awareness for its proper utilisation. That’s where the role of the open-source community comes in, wherein access to this knowledge is also facilitated,” says Chafe adding that the community can benefit only when these generated models are open-source, which he says is highly unlikely with MusicLM. One must not forget that ChatGPT Plus, a subscription-based plan, is also already here.
“But I don’t think this is avoidable at all at this point. For me as an artist, my best bet will be to learn to swim across the waves rather than fight it and drown. So, I’m just going to utilize it to my own purposes and for my own benefit,” Chafe concludes.
(The article was updated to makes changes to the headline on 10/02/2023 at 3.24 pm)
This post is released under a CC-BY-SA 4.0 license. Please feel free to republish on your site, with attribution and a link. Adaptation and rewriting, though allowed, should be true to the original.
Add a Comment