Clara Olivares en studio
In her new work, Au banquet des visages, for ensemble and electronics, which will premiere at the Grand Soir Numérique on January 8 at the Cité de la Musique – Philharmonie de Paris, as part of Némo – International Biennal of Digital Arts of the Île-de-France Region, and featuring the Ensemble intercontemporain, composer Clara Olivares focuses on the interaction between human and artificial intelligence. In particular, she invited the audience to help her constitute a database of vocal recordings.
At the sources of the Banquet: The need to get hands-on with AI”
The project Au banquet des visages originated from Clara Olivares’s desire to confront AI, often perceived as a contemporary societal “bogeyman”, particularly since the emergence of conversational models such as ChatGPT in 2022. Its sudden irruption into everyday life prompted in her a need to “engage hands-on with AI”, a motivation further reinforced by her reading of Harlan Ellison’s I Have No Mouth, and I Must Scream (1967). In this short story, an artificial intelligence takes revenge on its creators by inflicting suffering on the last survivors of humanity.
Olivares ultimately decided not to use this post-apocalyptic story as the basis for her work. While she has been focused for several years on processes of speech alteration and computerized processing of the voice, the idea of using AI as a tool for composition raised new questions and shed a different light on existing ones. First is the question of the artificial voice, an aspect that has been consistent in her work at least since her first chamber opera, Mary (2017), which features puppets and real-time voice processing. Next is the question of identity, and everything that it implies in terms of mirroring, doubling and metamorphosis, concepts which AI is currently deeply disrupting.
While the title is deliberately theatrical, evoking the most celebrated Last Suppers of the history of painting, the term “banquet” also refers to a space of coexistence, if not dialogue, as well as to the notion of “consumption, in this case, of human substance, as the ‘faces’ are in fact altered human voices, or fragmented identities, that never appear in full but only in pieces. Through this process, we are absorbing the living, leaving behind only the remains (the bones, the whimpers).”
The need for an ethical and respectful AI
Another crucial issue that comes up when using AI is ethical in nature and concerns the choice of a database to train the models. All artificial voices, and even more so when they are generated through deep-leaning processes, are actually based on real human voices. Yet most of the massive databases available to users are of dubious origin, all too often collected without any consideration for human or property rights.
Clara Olivares en studio à l'Ircam
“Since the very beginning, I had decided not dip into anonymous corpuses or questionable databases”, explains Clara Olivares. Instead, she turned to a quite innovative alternative, which also comes with its own issues. In addition to using a database of voice messages that she collected over the years from willing volunteers, Olivares also opened a call for contributions, inviting the public to send her voice messages in which they shared their dreams, wishes, or on the contrary their deepest fears.
“By doing this, I was both acknowledging the human dimension of voice synthesis and allowing the public to step up from their position as an audience and become an active part of the composition process.”
To remain perfectly “ethical” and respectful of the volunteers, Clara Olivares sought to preserve their anonymity by implementing a specific protocol for the use of these recordings, ensuring that no human ear could identify them. To do so, she first needed to “deconstruct them before artificially reconstructing them with added alterations”–a rather paradoxical process because, while striving to emphasize human thought and expression, it simultaneously strips away individual peculiarities and merges all the voice together.
First, the recordings were put through a ‘speech-to-text’ tool in order to transcribe their content, which was immediately encrypted. The encoded messages were then processed through an AI model capable of deciphering and paraphrasing them (thereby removing any specific intonation or speech component that might betray the speaker’s identity) and then concatenating them to generate a single, extended monologue. “Using generative AI at this stage of the process is a crack in the idealistic tracking protocol I had set for this project (because this AI model was trained with massive datasets, the sources of which I could not trace), but it was the only way to completely preserve the anonymity of the original recordings”, observes Olivares.
This generated monologue was then processed through an ad hoc voice synthesis model to create a single voice. Together with computer music designers Pierre Carré and Victor Bigand, and researchers from the Sound Analysis and Synthesis team directed by Axel Roebel, Clara Olivares developed a hybrid system: a model trained with the collected vocal data to generate synthetic voices, which then undergo extreme transformation (including stretching, transposition, and filtering) such as to preserve the original vocal recordings but so heavily altered that they are rendered unidentifiable. A final touch of cross-synthesis combines these voices with other non-vocal components. In the final result, the identity of the created voice constantly and organically evolves, flowing from one personality to another, while accurately maintaining the proportion of female and male voices in the original recordings–a statistic that was calculated from the fundamental frequency of each vocal sample, without ever needing to listen to them.
A fresh perspective on AI
As fascinating as this extensive work on artificial voice was, Clara Olivares encountered a paradox: the more efficient the tool, the more realistic the result becomes, whereas for the composer, the beauty of this protocol resides in “the failures, the accidents, the unexpected deformation, the missteps. Those are the things that spark my creativity and that I can build on.”
Regarding AI as a whole, this experience made her even more pessimistic than she already was, particularly concerning the increasing role AI models are assuming in our daily lives, smoothing out individual speech peculiarities and homogenizing our modes of communication under the guise of efficiency. Clara Olivares concludes: “Using AI, however, made me focus even more on humanity, its slowness and fragility. A hundred years ago, Paul Valéry declared ‘Time past is now beginning.’ It is up to us now to inhabit both this time and this past world, embracing what makes us human and the interconnections that define our existence.”
Photo 1 : de gauche à droite , Pierre Carré, Victor Bigand et Clara Olivares
Biography
