Nvidia revealed an artificial intelligence program today at its annual InterSpeech conference that is superior to existing algorithms at handling intonation. The appearance of computer-controlled speech should be more humanistic.

Using generic adversarial networks, the research is quite similar to Nvidia's highly effective method of producing human faces (and random other objects) from data points of existing faces, which has been extremely successful. Nvidia's GPU Technology Conference (GTC) in 2017 also saw the introduction of an artificial intelligence voice for storytelling, albeit there were still some areas for development. Even though Nvidia released an enhanced version of the model in 2020 known as the Flowtron, this model was not capable of being actively updated when it made mistakes. With the new model, this is a possibility. A human voice actor can be guided in the same way that an artificial intelligence voice can be guided, according to the researchers. The spoken information is transferred to the AI model, which has been pre-programmed with the appropriate variables.

The artificial voice genuinely resembles the 'source,' in the same way that humans learn to speak a foreign language. This enables the algorithm to highlight specific words, pronounce them with more or less emphasis, and speak in a louder or softer voice, among other features.

The AI voice can replicate lyrics, but you can also sing, assist persons with speech problems in communicating, pronounce text in games more naturally, and even design applications that allow gamers to converse with artificial intelligence characters. The rest of this week, Nvidia has scheduled a series of demos and workshops that will go deeper into the approaches created for the new artificially intelligent voice technology. 

Have a peek at the video, quite impressive stuff.

Nvidia Develops AI Algorithm to Improve Computer-Assisted Speech