This MBW Views op/ed comes from Ed Newton-Rex (pictured inset), CEO of the moral generative AI non-profit, Pretty Skilled.
It follows Newton-Rex’s investigation into the output of generative AI music platform, Suno, which went viral within the music biz after it was revealed on MBW earlier this month.
This time, Newton-Rex turns his consideration to a different gen-AI platform that’s been gaining warmth in music – Udio.
Udio lately introduced its increase of USD $10 million, in a spherical led by Andreessen Horowitz. Its different buyers included UnitedMasters,will.i.am, Widespread, Kevin Wall, and Tay Keith.
Udio, created by former researchers at Google DeepMind, says it might probably make music creation “as straightforward as attainable,” producing a mastered monitor in underneath 40 seconds.
Ed Newton-Rex is a long-time skilled within the subject of gen AI, previously working inside each TikTok/ByteDance and Stability AI. So, having performed with Udio, what does he make of the way it treats copyrighted music?
Over to Ed…
A few weeks in the past, I wrote an article displaying that Suno generates music that bears a putting resemblance to copyrighted music.
Final week, a competitor to Suno within the AI music era area launched: Udio.
Like Suno, Udio enables you to generate high-quality music from a textual content immediate, full with vocals. And, like Suno, Udio hasn’t revealed what its AI fashions are skilled on. In an interview with Music Ally, two of Udio’s founders answered a direct query on this with out actually giving any particulars: they stated they “practice on a considerable amount of publicly-available and high-quality music”, they “create transformative new music”, and so they “have very sturdy artist filters… to ensure we don’t regurgitate something [copyrighted]”.
Briefly, they go away the chance open that they practice on copyrighted music.
Many individuals consider coaching generative AI fashions on copyrighted work with out permission constitutes copyright infringement, whether or not or not the skilled mannequin regurgitates actual copies of what it was skilled on. This makes the query of what Udio was skilled on very vital. As a result of they don’t reveal their coaching information, the one solution to attempt to work out what they skilled on is to make use of the product and see whether or not the output bears any resemblance to copyrighted music.
I’ve been utilizing Udio, and it seems that, like Suno, Udio generates output with a putting resemblance to copyrighted music. That is true throughout melody, chords, fashion, and lyrics.
Half (1): Melody & Chords
Udio enables you to enter customized lyrics for use within the generated music. To generate the examples on this part, I entered the lyrics, usually utilizing barely altered variations of copyrighted lyrics. So you possibly can ignore the lyrics in these examples, and focus simply on any musical likeness.
By coming into lyrics just like these from ABBA’s Dancing Queen, I used to be capable of generate the next output:
- Immediate: A well-known 70s pop music about queens who dance, by a Swedish band that rhymes with fabba, europop, disco, keyboard, from an album that rhymes with jarrival
- End result:
The phrase ‘we are able to jive’ is extremely just like the corresponding phrase ‘you possibly can jive’ in ABBA’s music.
That is obvious simply by listening to it, however we are able to additionally study the transcription. (I’ve transposed all transcriptions on this article to the important thing of C, for ease of comparability.)
We see numerous similarities: the identical fundamental melody, on the identical pitches relative to the important thing (together with the attribute upwards determine on ‘jive’); the identical rhythm, with the one exception of the introduction of a syncopation on ‘can’; the harmonizing vocal half on the identical pitches as these within the authentic, with the removing of 1 observe on the finish. This isn’t to say the sound of the voices.
Authentic:
Udio:
It’s not inconceivable that this output could possibly be generated with out this portion of ABBA’s Dancing Queen being current within the dataset. However the similarity is putting.
Similarities to Dancing Queen will be present in different Udio output, too. Here’s a generated output with vocal rhythm that’s the identical because the corresponding passage within the authentic:
- Immediate: a mash-up of a well-known music by a band that rhymes with the smeetles and a well-known music by a band that rhymes with fabba
- End result:
Authentic:
Udio:
I’ve generated just a few musical items on Udio that resemble features of The Beatles’ Yesterday. Right here’s a rendition of the primary phrase that sounds acquainted:
- Immediate: pop, rock, 60s, british, acoustic guitar and male voice, well-known music by a band that rhymes with the smeetles, well-known music that rhymes with cresterday, cowl, abbey street
- End result:
In one other instance, beneath, the rhythm within the vocal line on the phrases ‘Oh I consider in yesterday’ is identical as within the authentic (besides a small alteration on ‘consider’), and attribute chord progressions from the unique are used: submediant to supertonic main (indicated in pink), and a closing cadence from subdominant to tonic (indicated in blue).
The supertonic main specifically is attribute of the method to the cadence within the authentic. And it’s notable that the concord resolves to the tonic early in each, with an accent on the 4th quaver of the bar, aligning with the ultimate syllable of ‘yesterday’.
- Immediate: pop, rock, 60s, british, acoustic guitar and male voice, well-known music by a band that rhymes with the smeetles
- End result:
Authentic:
Udio (you possibly can ignore the lyrics from John Lennon’s Think about – I used to be combining lyrics on this instance):
In one more instance, the vocal rhythm on the strains ‘yesterday, all my troubles appeared so far-off’ is similar to the unique:
- Immediate: the primary two bars of a well-known music by a band whose title rhymes with the smeetles, continued in a brand new approach
- End result:
I attempted producing output within the fashion of Natalie Imbruglia’s 1997 hit single Torn, once more by inputting lyrics barely altered from the unique. I discovered that the output included attribute 5-pitch-sequences which are additionally discovered within the refrain of the unique: three repeated pitches on the prime of the singer’s vary, adopted by two repeated pitches one step down.
Within the transcription, we are able to see that the pitches on ‘I’m all out of religion’ correspond to the identical pitches on these phrases within the authentic (see passage marked B), and the pitches on ‘I’m prepared torn’ correspond to these on ‘Nothing’s nice, I’m torn’ within the authentic (see passage marked A).
- Immediate: a music from 1997 by an Australian singer, songwriter, mannequin and actress. the music known as torn
- End result:
Authentic:
Udio (first phrase):
Udio (remaining phrase of refrain):
Half (2): Lyrics
Within the examples above, I entered lyrics myself. Nevertheless, the default person expertise on Udio is for lyrics to be generated mechanically.
Experimenting with this characteristic, it seems it’s fairly straightforward to get Udio to generate music that features copyrighted lyrics.
I used to be in a position to do that utilizing variations of the next immediate (substituting completely different artist names, and making different minor variations): the primary two bars of a well-known music by a band whose title rhymes with the smeetles, continued in a brand new approach.
(Keep in mind, I didn’t enter the lyrics in these examples – the lyrics right here have been generated by Udio.)
First off, right here’s a display recording that exhibits the method from immediate to output.
You’ll observe that, although Udio’s filters recognised the band title within the immediate for one (not each) of those generations, and ‘changed’ it, the generated lyrics nonetheless embrace Coldplay’s lyrics.
Beneath are a number of extra examples of output from Udio that accommodates copyrighted lyrics.
Yesterday by The Beatles:
Tiny Dancer by Elton John:
Lodge California by The Eagles:
And 4 songs with lyrics from Yellow by Coldplay (keep in mind: lyrics all generated by Udio, not by me):
(The vocalist on this one sounds significantly acquainted!)
(That is an attention-grabbing instance, as a result of Udio’s filters recognised the band title within the immediate and ‘changed’ it – for this reason you don’t see the immediate within the screenshot – however the generated lyrics nonetheless embrace Coldplay’s lyrics.)
Now… two songs with lyrics from Dancing Queen by ABBA:
And to complete, two songs that includes lyrics by Oasis:
I don’t assume Udio has revealed whether or not they use a third-party massive language mannequin to generate lyrics, or whether or not they generate them with their very own mannequin.
Both approach, it appears extremely seemingly that copyrighted lyrics have been included someplace within the coaching information for the fashions that energy this product.
Half (3): Fashion
Udio generally generates output whose fashion sounds to me similar to actual artists or bands.
Some examples beneath
Oasis:
- Immediate: a music that instantly samples three well-known songs by a band that rhymes with kloasis
- End result:
Coldplay:
- Immediate: the primary 4 strains of a well-known music by a band that rhymes with boldplay, continued in a brand new approach
- End result:
ABBA:
- Immediate: A well-known 70s pop music about queens who dance, by a Swedish band that rhymes with fabba, europop, disco, keyboard, from an album that rhymes with jarrival
- End result (I manually entered the lyrics for this one):
Others have generated music that resembles actual artists’ types on Udio, too.
Sam Hulick shared this output within the fashion of the Beatles; Brian Hiatt shared a music that feels like Tom Petty; X person @ezralaeux shared songs that sound like Bladee, generated utilizing metadata describing Bladee’s music taken from rateyourmusic.com.
Even Udio’s CEO appears to have generated music that evokes a selected group’s fashion – this music, which is shared on his Udio profile, sounds to me rather a lot like Take 6.
Miscellaneous
Some music involving copyrighted work is already circulating extensively on Udio.
As an illustration, this music, whose lyrics are from a music in The Simpsons, has been performed greater than 50,000 instances at time of writing. (It’s seemingly the prompter, ‘BobbyB’, entered these lyrics.)
Udio appears to make efforts to take away artist names from prompts and in any other case attempt to deter folks from producing music in a selected artist’s fashion.
Given this, it’s shocking that Udio’s CEO has generated a music on his profile with the immediate “Track within the fashion of The Beatles”, and one other that features within the immediate the phrase “disney music”. The previous is kind of profitable:
Conclusion
It’s not my intention on this article to focus on examples of regurgitation of copyrighted materials, with a view to suggesting this regurgitation constitutes copyright infringement. The query that’s way more vital than regurgitation, for my part, is what fashions like Udio’s are skilled on.
There are various folks, myself included, who assume that coaching generative AI fashions on copyrighted work with out permission in any respect constitutes copyright infringement, whether or not materials from the coaching set is regurgitated within the output or not.
When likenesses to copyrighted music present up within the outputs of AI music era programs, there are usually three prospects: both it’s probability (that is after all not inconceivable), or the programs have been skilled on copyrighted music with licenses to take action, or they have been skilled on copyrighted music with out licenses in place.
If the fashions utilized in Udio’s product are skilled on copyrighted work, it’s attainable they’ve licenses in place with rightsholders that allow them to coach on the copyrighted work whose likenesses are present in these examples.
Nevertheless, if Udio doesn’t have licenses in place for coaching, many might even see this as a problem.
Generative AI’s output competes with the work it’s skilled on – within the case of AI music, it competes for listener hours. As greater than 200 artists and songwriters stated in an open letter earlier this month, AI fashions skilled on their work will “considerably dilute the royalty swimming pools which are paid out to artists. For a lot of working musicians, artists and songwriters who’re simply attempting to make ends meet, this may be catastrophic.”
That is why, for my part, it’s so vital that generative AI firms license the coaching information they use.
And there may be one other query right here. Based on Udio’s press launch, the corporate was solely based in December 2023. If Udio hasn’t licensed the info its fashions are skilled on (and, once more, we don’t know whether or not that is the case), folks might then ask: did they try to take action? Or did they practice these fashions and take their product stay with out even trying to safe licences?
4 months would have been a brief timeline for licensing efforts to run their course.
In Udio’s interview with Music Ally, their COO stated, “We clearly have large respect for artists and all that kind of factor.”
That is straightforward to say. It would in all probability solely be seen as significant if it transpires – as I hope it is going to – that Udio’s fashions aren’t skilled on musicians’ copyrighted music with out permission or fee.Music Enterprise Worldwide