LARRUZ tests MusicGen, an AI trained on over 5,000 songs, against Google's MusicLM. The open-source model can generate music from text prompts and remix existing tracks, though it struggles with longer pieces and some specific requests.
Questions this video answers
More than 5,000 songs (speaker statement).
30 seconds (speaker statement).
Two minutes (speaker statement).
Published in January of the same year as the video (speaker statement).
Chapters
Yes. This new AI has listened to more than 5,000 songs. And what did it learn? Well, it learned to perform text to music, similarly to Google's MusicLM. But this technique is truly something else. This is MusicGen, an AI where we enter the kind of music we wish to hear and it just gives it to us like this. Wow. Yes. All this was generated by an AI. This was a pop dance track with catchy melodies, tropical percussions and upbeat rhythms with the new Music gen. And now let's compare it to Google's Music lm. This one is real good too, and it was published in this year's January. However, the first one you've heard, musicgen has something up the sleeve. It is an open source technique that is free for everyone. Wow. Now, smooth jazz with a saxophone solo, piano chords and snare drums. First with MusicGen. That was real good. With a bit of a saxophone solo, perhaps a snippet. Now the previous work Music lm. I have to say that I really liked the smooth jazz part better here. That was real good. However, no saxophone. It is one thing to do well, but it is not enough. It also has to follow the prompt as well. Now wait a minute. These samples are 30 seconds long, but real songs are typically much longer than that. So is that it? No real songs for us? Well, don't despair, because this paper has two more absolutely incredible features. One, Long Music generation. Oh yes, these are two minute long samples. But we are experienced fellow scholars here, so we are looking for long term coherence. That means that we will look at snippets of different parts of the song and listen if they really sound like they belong to the same song. Note that the previous Music LM was not as good in this task. Here is a snippet we looked at in a previous episode. And now let's see if the new Music gen is any better on this. I can't wait. Let's do it. Wow. First one is Flying Colors. Not even a little bit of a drift. Second one, Maybe you noticed that I liked this one so much that I was a little reluctant to skip forward because I wanted to just keep listening. Coherence was not as tight as with the previous one. Art and music is subjective. But if we have a song that has a little more variation, this is a good one too. Loving it. And third, This absolutely knocked it out of the park. Absolutely. The same song for two minutes. This is super difficult to pull off. And in my opinion, the previous Music LM was not as good as this new technique. So much improvement in just a few months. Incredible. But we haven't looked at my absolute favorite feature yet. I am pretty sure it will be yours too. They call this feature conditioning. This essentially can take an already existing piece of music and remix it with a new text prompt. Listen to the input. And the output. Wow. This is insanity. However, not even this work is perfect. Here is something that you might call somewhat of a failure case. Yes, the input was maybe recognizable, but it was jumbled up and not nearly as good as the other samples. But wait, here comes the best part. All this is available to all of you free of charge. The source code of this paper is available and as of the making of this video, you can even run it on the web. I've put a link to it in the video description. So in the future we will all have our own AIs that create a nearly infinite amount of music personalized to you in real time. And just imagine what we will be able to do just a couple more papers down the line. What a time to be alive. What you see here is a report on this exact paper or topic we have talked about which was made by Weights and Biases. I put a link to it in the description. Make sure to have a look. I think it helps you understand this paper better. Their system is designed to save you a ton of time and money and it is actively used in projects at prestigious labs such as OpenAI, Toyota Research, GitHub and more. And the best part is that Weights and Biases is free for all individuals, academics and open source projects. It really is as good as it gets. Make sure to Visit them through WNB ME2MP or just click the link in the video description and you can get a free demo today. Our thanks to WeightsAndBiases for their long standing support and for helping us make better videos for you.
