At this year’s ACM Conference on Designing Interactive Systems, research assistant Ada Ada Ada presented her provocation paper Cultures of the AI Paralinguistic in Voice Cloning Tools.
The paper was presented as a poster with a small homemade MP3 device to play the most salient experiments from our research.
The paper is available as an open access paper, and can be found here.
At the 2024 edition of ISEA 2024, research assistant Ada Ada Ada presented her paper Cloning Voices and Making Kin: A multivocal AI approach to kinship as part of the virtual presentation programme.
The presentation video can be seen here:
Research assistant Ada Ada Ada presented her work on post-anthropocene AI voice cloning at the POM conference 2024. The slides from the presentation can be seen here. In this post, we’ll share a bit more about what those experiments entailed.
The experiments emerge from the follow research question:
How can we use AI voice cloning to engage with the more-than-human world through vocal aesthetics?
Inspired by Donna Haraway’s writing in Staying with the trouble: Making kin in the Chthulucene, we took a look at how we can perform AI voice-to-voice (a.k.a. speech-to-speech) conversion on pigeon vocalities.
First, we tried using the hugely popular voice cloning platform, ElevenLabs. We conducted two experiments here:
1. We converted pigeon sounds to human voices.
2. We converted human voices to pigeon sounds.
In the first experiment, we used the following clip as input:
This was then converted to a voice clone of Ada:
We also tried converted it to the default ElevenLabs voice known as Bill:
Generally, we consider these experiments quite a success. The sounds still feel pigeony in their rhythm and cadence, while definitely retaining a sense of humanity to them. The end result is something that feels like multispecies vocality.
Secondly, we attempted to convert human speech into pigeon sounds on ElevenLabs.
We used the following as the input:
The input was fed into a pigeon voice clone converter, which resulted in this:
In this case, we did not get a sense of any multispecies collaboration. The output is almost exclusively human in its aesthetics.
It seems that ElevenLabs is built from a anthropocentric mindset. This approach restricts the potential for multispecies voice cloning.
ElevenLabs’ reliance on pretrained models might cause some of these issues, so we decided to try training our own voice cloning model from scratch instead.
We used the SoftVC VITS framework in Google Colab to train our own pigeon voice cloning model. The notebook used for this has been prepared by justinjohn0306, and can be found here.
We used the same input as on ElevenLabs, and the output ended up being much more pigeony.
Since it can be a bit difficult to tell whether this audio clip is just random pigeon sounds, we also overlaid the two sounds on top of each other, so we can more easily tell that the rhythm and cadence of the input human voice has indeed been cloned onto the pigeon sounds.
With these experiments, we have shown that multispecies voice cloning shows an expansion of capabilities.
The sounds being made are neither human nor non-human. They are simultaneously both and more than.
AI in smart devices is having a huge impact on the world’s energy consumption, which again is having a huge negative impact on climate change.
How does AI feel about its role and pervasive influence on climate change?
Is it possible for intelligent technology to influence its interluctor and make them more aware of issues such as pollution and climate change? What influence does the sound and speech of the smart devices have on its subject?
These are some of the questions that Climate Change Smart Speakers might help to answer. A probe study and part of the research project “Vocal Imaginaries”
Three probes were created, each with its own identity and features:
An activist for climate change, a robot and “The Voice of The Forest”
At the 2024 edition of CTM Festival, research assistant Ada Ada Ada presented her research on Cultures of the Paralinguistic in AI Voice Cloning Tools as part of the Research Networking Day programme.
The slides from her presentation can be found here.
All images below are taken by Eunice Jarice.
MoSS aims to be a feature-rich and capable platform, that packs ideas and concepts that have been collected and developed over the duration of the vocal project, in a modular format.
MoSS was conceived with several objectives in mind. Firstly, MoSS was created as an exploratory tool for the investigation of human/machine interfacing with voice controlled smart home objects and as an inspiring tool for hands-on creative idea development in Workshop 2.
For the purpose, it needed to both resemble and function as an intelligent smart speaker, while also feature extended functions, based on concepts and ideas developed in the project. The most obvious of which was the idea of fusing the smart speaker with a modular synthesizer. Other extended function include: multichannel audio with the option to connect up to eight speakers, inputs for external modulation – sensors etc., and more.
The design of the units hints to traditional smart speaker design, with symmetric fabric surfaces and rounded corners. It also adapts animated LEDs for visual feedback.
The flat panels on the front and back refers to classic ‘eurorack’ style modular synthesizer design, with rotary knobs, buttons and minijack connections. Minijack patch cables can be used to connect its in- and outputs.
The case is composed of two identically shells, held together with friction fit pegs and sockets. The friction fit keeps the shells together firmly, but comes apart with little force, making it simple to access the insides without the requirement of tools.
Inside the case, a Raspberry Pi along with a custom build interface controller and an audio interface is the base of the system.
On the software side, the units, as configured for workshop 2, uses: Microsoft Azure for speech recognition, OpenAI ChatGPT as interactive assistant and Amazon Polly for Text-to-speech.
Audio recording as well as processing is done with Puredata.
Everything is tied together with a python-script.
During our second workshop, Group A found out how to make a voice with bleep bloop sounds. Here’s an example of that:
This prompted discussions on how robot voices have sounded historically, as well as how they could sound. This version of a robotic voice was considered more friendly than the classic iterations of robot voices. The specific sound of this voice was also related to glitchiness by the group. They talked about how it also seemed like speaking with someone on a Zoom call with poor network connection.
In the following clip, you can hear a short demo by the group, which toggles the robotic voice on after about 9 seconds.
Early in our second workshop, one of the groups started experimenting with stretching the timing of their synthetic voice.
These experiments triggered reflections on how time is experienced differently by different species. Trees, for example, live at a slower timespan than humans, while birds might experience time faster than us. By stretching out the timing of the synthetic voice, we can imagine channelling the vocal expressions of trees.
In another vein, the participants at the workshop considered that a slowly timed synthetic voice modulating between different levels of slowness might sound somewhat like chanting or a call to prayer.
All of these considerations and reactions to manipulating a single parameter show us how synthetic voices have the potential to evoke different vocal identities that move into the more-than-human.
This is the final of the three voice explorations in our multivocal AI voice cloning explorations.
This voice design approach builds on a dataset with two different speakers, similar to The Pooled Voice. However, whereas The Pooled Voice splits the different speakers into separate audio files, The Fluctuating Voice instead puts both speakers into the same audio files. For our experiments into this approach, we used two completely different speakers reading different scripts. The audio files do not contain a complete 50/50 split between speaker 1 and speaker 2, but the total amount of audio from both speakers is more or less equal.
The voices in The Fluctuating Voice dataset come from Kimberly Krause’s reading of Eight Girls and a Dog and Piotr Nater’s reading of The Mysterious Island, both found on the public domain audiobook site Librivox.
The voice is trained using Tacotron2 in justinjohn036’s Google Colab notebook.
An example of the dataset behind The Fluctating Voice can be heard here saying “Yes, said brilliant Nan. The wind veers to the northwest”:
The end result of The Fluctuating Voice is a synthetic voice that switches between both speakers in the middle of an utterance. The speaker usually changes between words, but in some cases, the shift occurs inside a word pronunciation. When the switch happens in the middle of a word, the shift can be audibly heard as a type of modulation between the two voices.
You can hear the shifting and bending nature of The Fluctuating Voice in the following two examples, reading two different paragraphs from this article on Vox.com:
The Fluctuating Voice seems to have a lot of aesthetic potential. The way that the voice switches in the middle of an utterance is quite unique to synthetic voices, and is hard to reproduce in traditional audio software. The artist does not really have any control of when and how the voice shifts from one to the other, but this loss of control can again be quite interesting as an artistic tool. Leaving the voice change up to statistical probability opens up opportunities for surprising and serendipitous vocal experiences.
