TED-RNN — Machine generated TED-Talks

Door: Samim.

“TED has become somewhat of an institution with the Silicon Valley intelligentsia. Over the years, they have accumulated 1904 speeches or 22.4MB of text which amounts to 4038409 words. All transcripts are readable here and cover a wide range of topics, from scientific to cultural.

While many of the speeches are fascinating and have captured the imagination of a generation, TED has generated its fair share of criticism. Benjamin Bratton referred to it as “placebo politics and mega-church infotainment” at his TEDx talk New Perspectives — What’s Wrong with TED?

I wrote a web-crawler in python that gathers the transcripts of all TED talks. This text-dump then got fed into the char-rnn library. After consulting with Andrej Karpathy, I experimented with a word-level fork of char-rnn which unfortunately yielded worse results compared to the character level model. A selection of the results then got fed to a text-to-speech synthesiser, creating three convincing TED-speakers which I named Jürgen TEDhuberAda LoveTED and Isaac TEDimov.