/i/2000820476.png?fit=%2C&ssl=1)
Nvidia presented a tool at the Interspeech 2021 conference with which AI voices can learn a natural pronunciation of words. Using the RAD-TTS tool, researchers can use a recording of their own voice to train a speech algorithm.
At the GPU Technology Conference in 2017, Nvidia researchers demonstrated the progress they had made in AI development. They also let out an artificial voice at the time, but were not yet completely satisfied with the performance.
In 2020, a new AI voice was presented: flowtron. This artificial voice sounded more natural and human, but the researchers still weren’t done. The next step, according to the researchers, was to adjust the algorithm when mistakes were made during pronunciation, in much the same way that it happens with humans: by means of imitation.
The researchers developed an AI model for this, called RAD-TTS, with which she a ai-text-to-speech-teach an algorithm how to pronounce a word, or group of words. They do this by uploading their own voice recording to the algorithm, converting it into parameters that can then be imitated by the algorithm.
With RAD-TTS, the pitch and sound of a recorded voice can also be drastically changed. This enabled one of the researchers to transform his own male voice into an artificial female voice. That voice was used as a voice-over in the promotional video. Some of the new technology is open source according to Nvidia and will be made available on Nvidia NeMo-toolkit.
–
–
/i/1373725292.png?resize=150%2C150&ssl=1)