“Created by Irish agency Rothco for the Times of London, ‘JFK Unsilenced’ used artificial intelligence to stitch Kennedy’s voice into the speech he had been set to deliver on the day he was killed in Dallas,” reports Alexandra Jardine. “Rothco and the Times worked with CereProc, a British audio tech company, to analyze recordings from over 800 Kennedy speeches and build a database from them. The 22-minute video appeared on the Times website.” To read the full article, click here.
In order to put the President’s voice to the 22-minutes of words he had prepared to deliver, with the help of tech company Cereproc, Rothco built the monologue entirely out of data, using AI. The speech was meticulously pieced together by analyzing analog recordings. The audio was then finessed at postproduction company Screen Scene.
The process involved isolating 41 phonemes for American English (the sounds that can be used to make any word) and stitching the small units of speech back together; reviewing 831 analog recordings of JFK speeches and interviews; analyzing speeches for noise, crosstalk and underlying sample rate; applying audio processing to remove noise; and applying spectrum analysis tools to improve acoustic environment and match them across samples.
The sounds were then carefully analyzed for pitch, energy, and spectral features. AI techniques such as Deep Neural Networks (DNNs) were applied to learn a model of JFK’s audio, and how he uses intonation in his speech. At synthesis time, DNN models and high-dimensional search were used to find the best set of units to create new speech audio. Finally, the speech was enhanced by months of painstaking sound engineering—allowing audiences to hear the Trade Mart Speech delivered in JFK’s own voice.
— The Times (@thetimes) March 16, 2018