A love song in French co-created with Artificial Intelligence.
role
Singer, song-writer, guitarist, music generation, text-to-image generation
authors
Ninon Lizé Masclef, Tomasz Sroczynski
prize
AI Song Contest
keywords
music generation, AI art, contest
I collaborated with the violinist Tomasz Sroczynski and Artificial Intelligence to compose a love song in French, entitled "Amour Stochastique". Our team Machine Forgetting has been selected as a finalist for AI Song Contest 2022.

Machine Forgetting · Amour Stochastique

In Amour Stochastique, we wanted to combine our contemporary culture with ancient traditions, mixing together various musical registers and historical temporalities into a modern love song in French. Privileging the aesthetic exploration to the historical accuracy, we intended at creating a patchwork of our own influences: folk, pop, medieval, psychedelic rock, bardcore. In the piece, humans and AI build together a second literacy reflecting upon the meaning of their cooperation on a love song. Where does the desire lie in the co-individuation of humans and machines ? The double meaning of the lyrics suggests AI is expressing feelings about itself, comparing the training phase to a longing for love. By drawing the analogy between love and creation, that are not optimization processes, we question the fear of incertitude and illusion of the real contingency of life offered by stochastic processes used in creative AI paradigms.

We started the composition by generating multiple audio samples with VQ-VAE (OpenAI Jukebox). We conditioned the generation on a violin sample recorded by Tomasz, multiple artists (Malicorne, Dead Can Dance and Estampie) and genres (medieval, folk, french, psychedelic). One sample caught our attention, polyphonic with voices singing in an unidentified language. It reminded us of traditional chants. From this single sample, we created several elements of the song, including the vocals line, verse-chorus structure, triplet rhythm and chord progression, on which we iterated throughout the process.
Parallely, we generated lyrics with Cedille (fr-boris-8bit) fine-tuned on love and philosophical quotes in French. Tomasz improvised a violin accompaniment, while Ninon recorded vocals, and an electric guitar that follows the harmony of the generated sample.
Since we believe it is important in a co-creative process that non-coders of the team experiment with AI tools, we also used pretrained models with GUI, e.g. MelodyRNN embedded in Magenta Studio, to create melodic lines from MIDI notations of music from Guillaume de Machaut. We transposed and sampled the output to make a riff for flute and harp VST. Finally, we trained several models (DDSP, RAVE) to learn representations of the sound of lute and polyphonic bardcore music. We kept the first model to turn the violin timbre into lute, adding local variations of expressivity coming from the improvisational practice of Tomasz.


The Machine Forgetting project was studied at Berklee College of Music as part of the Machine Learning for Musicians course. Below is the presentation that Dylan Ever gave at Berklee about my project: