Canadian AI Startup Lyrebird Develops Imitation Algorithm That Could Mimic Any Voice

Lyrebird, an artificial intelligence startup in Montreal, Canada, developed an algorithm that imitates voice. To do that, the AI analyzes for a few second pre-recorded audio. The algorithm recognizes patterns in the speech of a person and reproduces the patterns through simulated speech.

The startup trains its models on a very large dataset with thousands of speakers. Lyrebird then compresses the information for a new speaker in a small key with their voice DNA, Gizmodo reports. The company uses the key to say new sentences, Jose Sotelo, a speech synthesis expert at Lyrebird, explains.

As proof of its ability to fake any voice, Lyrebird made samples of American celebrities, namely former U.S. President Barack Obama, former State Secretary Hillary Clinton, and current U.S. President Donald Trump. Sotelo says the startup uses artificial neural networks that function similar to how the human brain’s biological neural networks operate. Although the samples still show digital artifacts, problems with clarity and other weird characteristics, listeners could easily identify who is the person being imitated.

Lyrebird is not the first to tap AI to create synthesized voices that appear like spoken by humans. The Verge reports that Adobe has Project VoCo which edits human speck similar to how PhotoShop digitally alters images. However, to mimic human voice, the Adobe prototype software needs at least 20 minutes of sample audio while the Canadian startup requires only 60 seconds.

In a press release, Lyrebird says the AI applications would offer to companies and developers new speech synthesis solutions. By using the voice-like algorithm, users could generate a whole dialogue by tapping a voice of their choice or create new and unique voices from scratch that would fit their needs. The company uses deep learning models developed by researchers at the University of Montreal’s Montreal Learning Algorithm Institute lab. Three of the lab’s founders – Alexandre de Brebisson, Sotelo, and Kundan Kumar – are current Ph.D. students.