Meta, formerly known as Facebook, has introduced its latest language learning model (LLM) called MusicGen, developed by their Audiocraft team. This AI-powered tool allows users to input text descriptions of the desired music style, and MusicGen generates a 12-second track accordingly. While MusicGen has its limitations, it offers a glimpse into the potential of AI-generated music and its impact on crafting soundtracks and tunes. However, challenges such as copyright issues and the tool’s current capabilities need to be considered.
Meta’s MusicGen is making waves in the world of AI-generated music. Developed by the company’s Audiocraft team, this LLM functions similarly to ChatGPT but with a focus on music composition. Users can provide a brief text description of the desired music, click the “Generate” button, and MusicGen creates a 12-second track based on those instructions. For example, requesting a “lofi slow BPM electro chill [song] with organic samples” results in a track reminiscent of the popular Lofi Girl radio on YouTube.
What sets MusicGen apart is its ability to be “steered” by uploading existing songs. This feature allows the AI to gain a better understanding of song structures. Felix Kreuk, one of the developers behind the LLM, shared samples on Twitter showcasing how MusicGen can take classical compositions, like Sebastian Bach’s Toccata and Fugue in D Minor, and add drum beats and 1980s-style synths to create a more upbeat rendition.
We present MusicGen: A simple and controllable music generation model. MusicGen can be prompted by both text and melody.
We release code (MIT) and models (CC-BY NC) for open research, reproducibility, and for the music community: https://t.co/OkYjL4xDN7 pic.twitter.com/h1l4LGzYgf— Felix Kreuk (@FelixKreuk) June 9, 2023
Currently available to the public on Meta’s Hugging Face website, MusicGen has some limitations. Unlike Google’s MusicLM, which includes vocals, MusicGen focuses solely on instrumentals. This limitation may be a positive aspect since MusicLM’s vocal outputs often resemble unintelligible Simlish. However, musicians need not worry about AI replacing their careers just yet. While MusicGen is capable of creating simple and short melodies, its quality falls short compared to human-made compositions. The generated songs can become repetitive as MusicGen cycles through the same progressions multiple times. It is more suitable for creating background audio for videos or presentations rather than producing engaging music. The next pop hit is unlikely to come from an AI-generated source—at least for now.
For those interested in trying out MusicGen, it is advisable to act swiftly due to a couple of reasons. Firstly, the Hugging Face website’s stability is questionable, with reports of frequent crashes and dead links, potentially caused by high user traffic. Secondly, Meta may face legal challenges due to their use of licensed music tracks during the training of the LLM. Artists have recently filed lawsuits against developers and platforms for alleged copyright infringement. Meta might find itself in the crosshairs of litigious musicians or face industry pressure from record companies attempting to protect copyrighted content.
Meta’s MusicGen represents a significant step in the realm of AI-generated music. While it offers creative possibilities for crafting soundtracks and tunes, it currently falls short of human ingenuity in terms of quality and complexity. Its instrumental focus and limited capabilities make it more suitable for generating simple background audio rather than engaging musical compositions. Additionally, potential legal challenges related to copyright infringement pose a risk for Meta, considering their use of licensed music during the training of the LLM. As AI continues to advance, it remains to be seen how it will shape the music industry and whether it will coexist or compete with human musicians.
