AI can even come to the rescue of History. Our everyday colleagues The worldin partnership with the Institute for Acoustic/Music Research and Coordination (Ircam) and its technological subsidiary Ircam Amplify, was able to reconstruct the call of June 18, 1940, as it would have been pronounced by General De Gaulle, inviting in the hours the darkest of all the French not to give up the fight against the German enemy.
An audio corpus from the BBC
Never recorded, this call has often been reinterpreted and the precise text is still a source of controversy among historians. To obtain the most accurate version possible, the team relied on documents “based on Swiss sources, the British and French press and historical research on the subject”assures Charles-Henry Groult, head of the video service of the evening daily in a question and answer session online.
To reconstruct the original voice of Charles De Gaulle, as it was most likely heard on June 18, 1940 when it was broadcast, the Ircam Amplify researchers relied on several later speeches, recorded on the BBC by the general between June 22, 1940 and June 6, 1944. The sounds were then analyzed by an artificial intelligence which was able to recreate similar sounds. “The system extracts what makes the personality of the general’s voice, the color of his voice, his timbre”, explains Frédéric Amadu, technical manager at Ircam Amplify. A sound personality which will then be affixed to the sentences pronounced by an actor, here François Morel.
reproduce the emotion
Named voice cloning, this technique is extremely effective in reproducing a person’s voice. In total, the reconstruction required about a month of work, “including the recording of the actor, a full week of machine learning and the intervention of researchers, technicians and sound engineers to correct the audio”. To produce audio almost identical to the original speech, the Ircam model still requires numerous human corrections. In addition, the program used by the researchers is distinguished by its vocal precision, even in the reproduction of emotion that publicly available software is not yet able to replicate.
To reproduce historical voices, Ircam must respect a strict internal protocol aimed at guaranteeing the ethics of the projects carried out. “What is pronounced by the synthetic voice must have already been actually pronounced or written”, explains Frédéric Amadu. “What is the use of similar tools open to all: supervised? Respectful? Open ? Manipulator?”he asks himself.