If you have played ChatGPT, you are aware of how impressive (and dangerous) AI can be. ChatGPT can generate virtually whatever you request, whether you provide basic or sophisticated instructions. Google has found out how to accomplish the same feat. However, instead of producing poems or computer programmes, Google’s AI composes music, and you can see its works right now.
Google’s artificial intelligence, codenamed MusicLM, is supposed to generate many minutes of high-quality music based on text cues. While artificial intelligence music is not necessarily novel, the company says that its algorithm generates music that is more true to the initial text prompt and of higher quality than earlier models. Undoubtedly, they are bold assertions. However, based on these previews, they may exceed expectations.
The initial batch demonstrates how MusicLM can generate music from rich captions, similar to how OpenAI’s ChatGPT generates its renowned user answers. Google can provide MusicLM with a caption such as “The primary soundtrack of an arcade game.” It is fast-paced and lively, with an electric guitar riff that is catchy. The music is repetitious and easy to recall, but contains surprising sounds such as cymbal crashes,” and the AI will process a recording that matches these characteristics.
The music developed from “Funky piece with a powerful, danceable beat and a noticeable bassline” is my particular favourite in this segment. On the second page, “A captivating keyboard melody adds depth and complexity to the tune,” which sounds like something from Stardew Valley.
When Google asks MusicLM to make vocals, though, the uncanny valley really begins to set in. Many of the examples on this page contain AI-generated voices, and while some of them seem “computer-like,” others sound a bit too human. It is somewhat unnerving to listen with the awareness that the voice is entirely synthetic. Google even requested MusicLM to make a rap song, and the “rapping” seems disturbingly genuine, despite the fact that none of the words are actual.
If you want to be truly frightened, scroll down to “Text and Melody Conditioning.” Here, you can hear a variety of voices and instruments attempting their best with various songs, with wild outcomes. The first choice features MusicLM humming “Bella Ciao,” which sounds excessively wonderful. White Lotus is produced by the “Tribal drums and flute” version.
Also available is “Long Generation,” which generates five-minute-long tunes from brief descriptors such as “melodic techno” and “relaxing jazz.” However, “Story Mode” adds a new level of intrigue. Google feeds MusicML several captions, and the artificial intelligence modifies the song based on the caption. The caption begins with “time to meditate” for 15 seconds before changing to “time to wake up.” As though beginning a new verse, the music shifts effortlessly into a different key, accompanied by eerie vocals.
In spite of the creepy vocals, it is intriguing to observe the AI adapt the song without changing it. The majority of the time, the new caption is incorporated into the song without creating the impression of a completely different melody. Particularly amazing is the AI’s ability to go from “pop song” to “rock song” to “death metal song” in the third example. While “rock music” may not sound particularly rock-like, “death metal song” is absolutely wonderful. I believe AI Death will be a smash among metal fans.
Google conducted an interesting experiment in which it fed MusicLM descriptions of renowned artworks to determine what music it would generate. If you’ve ever wondered what “The Starry Night” by Vincent Van Gogh would sound like as a song, now is your chance. It is also fun to explore how MusicLM develops particular instrument and genre sounds. The electric guitar is excellent, and “British Indie Rock” features a dramatic shift near the conclusion of the preview.
In addition to the “Accordion Solos” and analysing different responses for the same prompt, there is more to discover in this MusicLM sample. Google has also produced a 15-page research paper on MusicLM if you are interested in reading about the system’s technical aspects. While I hope the corporation publishes a beta version for public consumption, these samples are the next best thing.