ShuhBang!

Generative Neural Networks or GNN's are currently really taking off. Most of them focus on natural language. These include LLAMA and GPT, and there are many other examples out there. There are also still image generators, and even video generators. Some very interesting aspects of the image and video generators are the use of a text prompt to help define the end goal. A team at Meta has developed a music generator. This application is called MusicGen. You can use it in a few ways. First, you can supply a text prompt, and it will build a song for you. Second, you can provide a text prompt with a song, and it will use the text prompt to alter the provided song to match the text prompt.

There are some issues that need to be taken into account. First is copyrights. GNN's need to be trained on curated data in order to learn. In this case, curated data are existing songs with several attached classifications. For example, they would take "Black Dog" by Led Zeppelin and attach labels to it, such as the song title, the artist, the genre: "Heavy Metal", the decade:"70's", "hard guitar", "Heavy drums". There are countless other labels they could attach. You need to purchase rights to the music in order to do these operations. 

The next issue is the hardware involved. The documentation for the project states a minimum 16GB GPU needs to be used in order to create songs in a relatively short time. I tried using this on a 48 core server with 750GB of RAM and it took 48 hours to generate a 30 second clip. A 16GB GPU is supposed to take a few minutes to generate a 3 minute song. I would say the more CUDA cores, the faster the generation. The main problem here is Bitcoin miners are still hogging all the new GPUs so the price for a high CUDA core, large memory GPU is still very high. A 24GB Nvidia 4090 is still $2,200CAN. A100's and H100's are twice that. 

The final issue is how you are permitted to use the application. Meta claims MusicGen is Open Source, but the model weights are licensed for non-commercial use, and only the code that provides an example of how to use of the model weights is Open Source.

We'll see how these things evolve.