Talk To Einstein

Project Goals:

• Develop novel use cases for synthetic media.

• Create engaging synthetic learning content.

• Expose viewers to potential AI applications in the humanities.

• Explore ethical dimensions of synthetic media creation.

Talk To Einstein is an experiment in algorithmically generated (or “synthetic”) media (AGM). I fine-tuned the OpenAI GPT-2 model for predicative text generation with several archival text correspondence collections, for example the complete papers of Albert Einstein, Charles Darwin, or selections of Civil War soldier correspondence from the the Library of Congress. Based on these large data sets of written letters, the AI model is able to produce novel text that takes the form of written correspondence, with style and content consistent with the data source. You can listen to a synthetic sample of Darwin’s “writing” here, or of a Civil War soldier here. Note in both examples, the generated text sounds plausible, but appears no where in the original correspondence.

For Albert Einstein, I collaborated with resemble.ai, a Toronto-based AI company specializing in synthetic voices, to locate and parse all known English language audio recordings of Einstein speaking, in order to create an AI model capable of reproducing a recognizable version of Einstein’s voice.

Here is an actual recording of Albert Einstein speaking:

And the same words, spoken by AI Einstein:

Here is a sample of AI Einstein speaking words written by Albert Einstein, but for which there is no known audio recording:

I have also incorporated synthetic video techniques to “animate” a single photograph of Einstein with my own video, to recreate a talking Einstein representation.

You can see an example of the process, using a single photo and a driving video below:

There are a great many compelling use cases for this general approach to synthetic media and the humanities, which I shall continue to explore with my students and colleagues. At the same time, there are nearly as many unsettling applications for potential misuse, equally in need of examination and careful thought.

Discovery Engine:

On of the more interesting applications for synthetic text generation to academic humanities is as a discovery engine for archival inquiry. Can our conversations with AI Einstein make new discoveries and connections within Einstein’s own archive? Natural language search has the possibility of greatly expanding access and interest in humanistic archival research.