Artificial intelligence mostly isn’t, at least not yet. Still, the artificial neural networks that humans have created are capable of some very impressive things, and you’ve probably seen plenty of examples of text and images created by computers that are extremely cool. What about music, though? There have been attempts to make AI music in the past, but for the most part they haven’t been great. Enter Google Research’s MusicLM.
MusicLM builds on earlier work and is, according to its creators, “a model generating high-fidelity music from text descriptions.” Indeed, in that sense, it is to music what Stable Diffusion or DALL-E are to images. Users provide text, which can be as little as a single word or as richly-descriptive as you want, and the AI model attempts to re-create the sounds you wanted to hear.
Besides working from text prompts, Google’s team demonstrates that MusicLM can also accept a hummed or whistled melody as embedded conditioning, which it can then reproduce using a variety of instruments. It’s fascinating to hear the AI’s take on a simple fingerstyle guitar solo as recreated by an opera singer or a string quartet.
While some of the results are compelling—we were particularly impressed by the AI’s take on Reggaeton music—many of them are of rather low quality and are missing elements described in the text caption. In that sense, it’s rather like other generative AIs where the quality of the results can vary greatly.
It’s still impressive how often MusicLM gets it right, and also the degree of specificity that you can use in the prompts. Somewhat-niche descriptions like “east coast hip hop” and “minimal house” seem to generate genre-accurate musical snippets with remarkable faithfulness to the original sound—even if the rap lyrics are garbled and unintelligible.
The Google Research website for the project has dozens of audio samples for your examination, and we listened to quite a few, but we recommend you to head over there and try out some for yourself. As Google’s team says, this is probably the best AI music generator to date, and things will only improve from here.
