'It was always the plan to put the world in your hand', or so Bo Burnham gibed in his lockdown musing on performativity and creativity. The latest wave of the digital revolution, Artificial Intelligence (AI), has been quietly colouring the world of song writing and music production for some time. How to introduce students to the ever-growing array of tools for music creation, while combating the omnipresent challenge of plagiarism, is a task that requires us to rethink traditional notions of authorship and creativity in the digital age.
Looking at software options is a good place to begin.
Create
Splice quietly released their new AI feature, Create, in Beta at the end of the summer, and it has been making subtle waves throughout the music production community. For anyone unfamiliar with the platform, Splice.com is a hub for music production, bringing together freemium sample libraries, loops, instruments and plugins, projects, and sounds from over two million artists to create an online community of music-makers. Create, the latest addition to Splice's impressive arsenal of creative tools, is their first foray into the murky world of AI.
Artificial Intelligence has become a tacit mainstay of traditional songwriting over the last few years, helping prompt rhymes, lyrics and even suggesting chord patterns. Attempts at completely generative music-making by AI have been wobbly at best. Create takes a new approach to generative music production. The engine matches samples from its extensive library to the style of your choice, suggesting combinations and balancing them to create an instant loop. The landing page invites the user to pick from a range of up to 20 pre-defined styles, or to select ‘no filter’.
The AI engine then combines four samples together. Where Create gets interesting is that the samples come from any part of Splice's two million-plus catalogue, rather than coming from the same sample packs or authors. This can lead to some unique and surprising combinations, or stacks. Rather than just searching for loops with the same tags – BPM, key, style, etc. – the engine applies some imagination to the combination, altering the key and BPM of the samples in order to fit them together harmoniously. The engine also cleverly combines loops of differing lengths with different loop points. This leads to an exponentially greater number of combinations than traditional sample matching, not to mention taking the legwork out of looking through thousands of samples for that one perfect sound which sparks the creative process.
Create's interface is beautifully intuitive; the initial combination of four or five samples can be soloed or muted individually just by clicking on the track. The project window allows you to combine up to eight samples at once. Empty slots can be filled with a single click, which prompts you to add Bass, Chords, Percussion, FX, Vocals or any instrument of your choosing. Samples can be discarded or replaced at the touch of a button and each sample has an independent volume fader, allowing for a proto-mix of the stack to be refined before settling on a final result. Finally, clicking ‘new stack’ repopulates the project with new samples according to the style and BPM prompts selected.
The engine also comes with a helpful feedback window, allowing users to comment on individual stacks.
Create: the road test
Having recently run a series of electronic production workshops, I decided to road test Create with the participants, as a tool for stimulating ideas and directions. Of the genres available, we found some to be more successful than others. ‘Retro Synth’ and ‘Soulful Beats’ produced the most consistent results; the more dance-based forms (Drum n Breaks and Disco Fever) produced some mixed results. Some of the stacks were useful jumping-off points; even the rejected ones provided a useful exercise in analysing what the participants did not like about the final result.
Create served as a useful framework for discussing ideas and aesthetic choices in production. This provided participants with starting points for their own blank projects, or new directions for tracks in progress. It also raised an interesting question about producing solely, or mainly, with premade samples and loops. While there is an undeniable creativity in combining loops in order to achieve a final result, the danger is that productions become generic. This is especially true where several producers jump on the same sample packs or loops. The artists who work effectively in this genre – Disclosure, Wax Tailor and Mau P, to name a few – are most notable for their imaginative approaches to incorporating samples. The genius of their music is in the manipulation and combination of their source material. Although Create is a powerful tool for navigating Splice's proliferating library, and a great spark for ideas and combinations like all of the AI songwriting tools, it is only as powerful as the musician using it.
Vocal Remover
Another of my AI-based discoveries over the summer has been VocalRemover.org, an AI-driven engine which uses a combination of phase cancellation and mirror EQ to separate out a track into the vocal and instrumental constituent ‘stems’. While this facility is not necessarily new, Vocal Remover is capable of achieving impressive results compared to its competitors. The engine is still in development and occasionally produces some funky results! But, like all AI, it has the ability to learn and improve based on user feedback. In the past, sampling existing tracks involved either begging record labels for access to the original stems, trying to separate out the parts manually (with mixed results!) or having to sample the final track, complete with the original instrumental accompaniment. This approach limits the scope of the remixer considerably.
AudioCraft
Another of the summer's big releases was Meta's AudioCraft. This includes three main tools: MusicGen, AudioGen, and EnCodec. MusicGen generates new music from text prompts or melody inputs; AudioGen works solely from text prompts; and EnCodec can improve the quality of generated audio. The suite can be installed via Google Colab or, for those comfortable with Python, via Anaconda. Unlike Create, AudioCraft allows the user to customise the duration of the track, however the platform offers less facility for customising the final result. EnCodec allows the user to adjust parameters in the audio created from the prompt, but the result is still a single stereo file, rather than a multitrack. This limits the scope for manipulating the result as a sample in the way we're used to.
The big advantage of this system is that it generates a unique starting point. However extensive the Splice, or similar, sample libraries grow, there is always the risk of producers using the same samples, taking away from the uniqueness of the final result. My production students found AudioCraft to be a useful starting point for producing the initial loops which sparked their projects. The ability to customise the output in terms of key, BPM, instrumentation, and style made AudioCraft a useful tool for production. Importantly, it places the emphasis on the producers’ own imagination, rather than on trawling through endless clips and loops, awaiting inspiration.
MusicLM
Not to be outdone, in July Google quietly released the Beta of MusicLM. MusicLM generates music from text prompts, melody inputs or even short audio samples. It can also generate music in a variety of different styles, including classical, pop, rock and jazz. At present, the engine has a waiting list, although it took two weeks for my invitation to come through. MusicLM is trained on a dataset of text and music, while AudioCraft is trained on sound effects and Meta-owned music. This means that MusicLM is better at generating music in a variety of genres, while AudioCraft produces sound effects which defy the imagination.
A word to the wise: legality. Although sampling has been one of the mainstays of commercial music for around 50 years, sampling ‘a substantial part’ of a track without the owner's permission can have serious consequences. My favourite example is Vanilla Ice's track ‘Ice, Ice, Baby’, for which the record label, SBK Records, did not clear the use of the famous Queen/Bowie bassline. Vanilla Ice was eventually forced to surrender half of the track's earnings to Queen and David Bowie. Always clear your samples before you let anything out into the world!
