![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
Generating Music in 2025 & a case for generating your own midi input datasets.
The technology deployed to consumer service
AI-generated music creates new compositions by analyzing existing compositions. This Ai technology can create compositions from audio samples or converted to text midi files. The Ai music generators build upon existing musical knowledge and structures to be found within its dataset. The quality of the output depends on the platform’s algorithms employed in analysis and the input data in the datasets it's been trained on.
From the consumers point of view
With just a few prompts from you, these Ai music generators produce original music that can be used in a number of applications, from live game play streaming, soundtracks for videos & online content creation. The reason to use these services if you need some music composed for the project. This is all about copywrite infringement filters on social media sites, taking down anything in copy right infringement a content creator or user posts that is in violation. Some AI music generators allow you to customize elements like mood, genre, tempo, and instrumentation, giving you a surprising degree of control over the final product. These changes can even be done on the fly selection to the Ai music generators, forcing the composition to change and recompose a new composition.
But is this really composing music from nothing and there are alot of questions that can be asked as to how each of these Ai music generating companies are manipulating the datasets? Which datasets are they using? Are they truly predicting an entirely new song from the datasets or just Ai Dj mixing the available songs in the dataset. In most cases these companies are deploying midi straight to the desktop. Performing the music on the MS GS Wavetable that comes standard with any Windows Desktop. A few of the companies have soundsets that you can install to improve the midi performed audio sonic qualities over those desktop sounds.
The audible results from the available company demos reminds me more of sound bubbles than constructed composition. There is a subtractive songlike quality to the Ai generated products. The melodies don't have structure if they exist at all. The results are soundscape, experimental. Which like vector and linear synthesis has been done. Yet the products produced from these Ai music generators is much more like subtractive synthesis. You start off with what I can only assume is a full high quality audio sample and you slowly slice away frequencies to whittle down the original to a lesser audio form of the original. Which is a safe way to go about using Ai to produce a like song for a customer today that used to sound like another song. Anything happening in these Ai employed algorithms that might remind us of FM or additive synthesis, where the end result has been built upon the original with additional frequencies? It would be great if these companies published descriptions of the algorithm functions employed in their models. In most cases the end results of these online generators fall somewhere into the game music category. From soundtracks to 8-bit vintage game music. Hard to imagine this audio being used in a professional performance.
I'm focusing on these types of Ai companies generating music,
https://sourceforge.net/software/product/Infinite-Album/alternatives
I will keep searching online to locate some technical papers on how these companies have built their Ai models, the algorithms employed in the model's frequency and composition analysis, a structured flowchart...anything that is more revealing on how the Ai algorithms are being used in generating the music. At this point I can find much. Trade secrets huh?
Fact number one is that these Ai music generators build upon existing musical knowledge and structures. So, there are three types of Ai music generators out there, those that use commercially available music and those that build on royalty free music and those that are doing what KicKRaTT is doing, generating their own midi input datasets. The aforementioned generators must have some form of input data or datasets to perform. The companies that are currently succeeding are either using the commercially available songs legally (paying royalty fees to the artists) or illegally (not-paying royalty fees to the artists). The difficulties I face as an artist using Ai to generate music I have expressed in this journal. There have been for a while, numerous ways to generate midi. Generating the input data for the datasets is the untapped original music that I focus on and there is so much to be had. No one artist could possibly tap into it. It's pretty ridiculous that so many people are running down this road of music technology in the other direction. When there is so much that can be founded on the many possible genres yet to be discovered from the many ways to generate midi score.
https://www.theverge.com/2024/6/24/24184710/riaa-ai-lawsuit-suno-udio-copyright-umg-sony-warner
The remaining music generating startups unable to pay the huge licensing fees that the record companies or RIAA are suing for turn to royalty free music datasets for Ai training.
Melodia appears to be at the center for a number of these Ai music generating startups. Music generators for content creators. A quick way to put music behind a video production or web project. After reading through a number of the Music Generating App companies on SourceForge (link above) I speculate that these music generating apps are either using an available online free royalty dataset or one that they have downloaded.
https://www.melodia.io/
Utilizing these royalty free music datasets these other Ai music generators are able to progress.
What would be the difference between these two dataset types? One filled with music made with exceptional instruments and recorded in the best studios, the other made of samples from numerous unknown sources. One filled with the midi compositions of music's greatest, the other filled with, well... everything else. There really is no choice if using Ai to generate music, a generated midi dataset is a given. You lose too much control in creating original material if you use a pre composed dataset. With a precomposed dataset you can only hope to achieve what you wanted. There is an element to which a musician molds a composition into structure. With a precomposed dataset the artist will fight the ability to mold the composition. Because you have no control over the dataset.
We can see where all of this going. The small Ai music generating companies are waiting for the lawsuits with the larger Ai music generating companies to pan out, fines to paid, agreements to be made and eventually licensing to become monetized. Put on a subscription basis for anyone to tap. So that in the end if the customer wants a song like Marvin Gaye and a song that the customer can call his own, use on his website, whatever. The customer will pay for it. Over time this cost will come down as companies level costs through other expenses to make their whole business model work. And in the end, it reduces the creative rights ownership of the original artist. This whole process will dramatically slow the creative output in real music, as we watch these companies demand that their music is heard and get into copywrite infringement battles.
You talk your way through enough of these blog posts and you uncover an entirely new number of reasons why if safeguards aren't put in place with Ai media art generation, in its many forms, Ai will reduce the creative value of the human artist. On the large side of mass consumerism, the argument will be made, "Why pay the artist when we can fabricate a number of variations just like the artist". In the general consumption of media, it seems a very possible outcome.
Final thought... be original, generated your own datasets and use Ai technology creatively.
The technology deployed to consumer service
AI-generated music creates new compositions by analyzing existing compositions. This Ai technology can create compositions from audio samples or converted to text midi files. The Ai music generators build upon existing musical knowledge and structures to be found within its dataset. The quality of the output depends on the platform’s algorithms employed in analysis and the input data in the datasets it's been trained on.
From the consumers point of view
With just a few prompts from you, these Ai music generators produce original music that can be used in a number of applications, from live game play streaming, soundtracks for videos & online content creation. The reason to use these services if you need some music composed for the project. This is all about copywrite infringement filters on social media sites, taking down anything in copy right infringement a content creator or user posts that is in violation. Some AI music generators allow you to customize elements like mood, genre, tempo, and instrumentation, giving you a surprising degree of control over the final product. These changes can even be done on the fly selection to the Ai music generators, forcing the composition to change and recompose a new composition.
But is this really composing music from nothing and there are alot of questions that can be asked as to how each of these Ai music generating companies are manipulating the datasets? Which datasets are they using? Are they truly predicting an entirely new song from the datasets or just Ai Dj mixing the available songs in the dataset. In most cases these companies are deploying midi straight to the desktop. Performing the music on the MS GS Wavetable that comes standard with any Windows Desktop. A few of the companies have soundsets that you can install to improve the midi performed audio sonic qualities over those desktop sounds.
The audible results from the available company demos reminds me more of sound bubbles than constructed composition. There is a subtractive songlike quality to the Ai generated products. The melodies don't have structure if they exist at all. The results are soundscape, experimental. Which like vector and linear synthesis has been done. Yet the products produced from these Ai music generators is much more like subtractive synthesis. You start off with what I can only assume is a full high quality audio sample and you slowly slice away frequencies to whittle down the original to a lesser audio form of the original. Which is a safe way to go about using Ai to produce a like song for a customer today that used to sound like another song. Anything happening in these Ai employed algorithms that might remind us of FM or additive synthesis, where the end result has been built upon the original with additional frequencies? It would be great if these companies published descriptions of the algorithm functions employed in their models. In most cases the end results of these online generators fall somewhere into the game music category. From soundtracks to 8-bit vintage game music. Hard to imagine this audio being used in a professional performance.
I'm focusing on these types of Ai companies generating music,
https://sourceforge.net/software/product/Infinite-Album/alternatives
I will keep searching online to locate some technical papers on how these companies have built their Ai models, the algorithms employed in the model's frequency and composition analysis, a structured flowchart...anything that is more revealing on how the Ai algorithms are being used in generating the music. At this point I can find much. Trade secrets huh?
Fact number one is that these Ai music generators build upon existing musical knowledge and structures. So, there are three types of Ai music generators out there, those that use commercially available music and those that build on royalty free music and those that are doing what KicKRaTT is doing, generating their own midi input datasets. The aforementioned generators must have some form of input data or datasets to perform. The companies that are currently succeeding are either using the commercially available songs legally (paying royalty fees to the artists) or illegally (not-paying royalty fees to the artists). The difficulties I face as an artist using Ai to generate music I have expressed in this journal. There have been for a while, numerous ways to generate midi. Generating the input data for the datasets is the untapped original music that I focus on and there is so much to be had. No one artist could possibly tap into it. It's pretty ridiculous that so many people are running down this road of music technology in the other direction. When there is so much that can be founded on the many possible genres yet to be discovered from the many ways to generate midi score.
https://www.theverge.com/2024/6/24/24184710/riaa-ai-lawsuit-suno-udio-copyright-umg-sony-warner
The remaining music generating startups unable to pay the huge licensing fees that the record companies or RIAA are suing for turn to royalty free music datasets for Ai training.
Melodia appears to be at the center for a number of these Ai music generating startups. Music generators for content creators. A quick way to put music behind a video production or web project. After reading through a number of the Music Generating App companies on SourceForge (link above) I speculate that these music generating apps are either using an available online free royalty dataset or one that they have downloaded.
https://www.melodia.io/
Utilizing these royalty free music datasets these other Ai music generators are able to progress.
What would be the difference between these two dataset types? One filled with music made with exceptional instruments and recorded in the best studios, the other made of samples from numerous unknown sources. One filled with the midi compositions of music's greatest, the other filled with, well... everything else. There really is no choice if using Ai to generate music, a generated midi dataset is a given. You lose too much control in creating original material if you use a pre composed dataset. With a precomposed dataset you can only hope to achieve what you wanted. There is an element to which a musician molds a composition into structure. With a precomposed dataset the artist will fight the ability to mold the composition. Because you have no control over the dataset.
We can see where all of this going. The small Ai music generating companies are waiting for the lawsuits with the larger Ai music generating companies to pan out, fines to paid, agreements to be made and eventually licensing to become monetized. Put on a subscription basis for anyone to tap. So that in the end if the customer wants a song like Marvin Gaye and a song that the customer can call his own, use on his website, whatever. The customer will pay for it. Over time this cost will come down as companies level costs through other expenses to make their whole business model work. And in the end, it reduces the creative rights ownership of the original artist. This whole process will dramatically slow the creative output in real music, as we watch these companies demand that their music is heard and get into copywrite infringement battles.
You talk your way through enough of these blog posts and you uncover an entirely new number of reasons why if safeguards aren't put in place with Ai media art generation, in its many forms, Ai will reduce the creative value of the human artist. On the large side of mass consumerism, the argument will be made, "Why pay the artist when we can fabricate a number of variations just like the artist". In the general consumption of media, it seems a very possible outcome.
Final thought... be original, generated your own datasets and use Ai technology creatively.
