The voices on Amazon’s Alexa, Google Assistant and different AI assistants are far forward of old-school GPS units, however they nonetheless lack the rhythms, intonation and different qualities that make speech sound, properly, human. NVIDIA has unveiled new analysis and instruments that may seize these pure speech qualities by letting you practice the AI system with your personal voice, the corporate introduced on the Interspeech 2021 convention. 

To enhance its AI voice synthesis, NVIDIA’s text-to-speech analysis workforce developed a mannequin referred to as RAD-TTS, a successful entry at an NAB broadcast conference competitors to develop probably the most sensible avatar. The system permits a person to coach a text-to-speech mannequin with their very own voice, together with the pacing, tonality, timbre and extra. 

Another RAD-TTS characteristic is voice conversion, which lets a person ship one speaker’s phrases utilizing one other particular person’s voice. That interface offers positive, frame-level management over a synthesized voice’s pitch, period and power. 

Using this expertise, NVIDIA’s researchers created extra conversational-sounding voice narration for its personal I Am AI video series utilizing synthesized reasonably than human voices. The purpose was to get the narration to match the tone and magnificence of the movies, one thing that hasn’t been completed properly in lots of AI narrated movies so far. The outcomes are nonetheless a bit robotic, however higher than any AI narration I’ve ever heard.

“With this interface, our video producer could record himself reading the video script, and then use the AI model to convert his speech into the female narrator’s voice. Using this baseline narration, the producer could then direct the AI like a voice actor — tweaking the synthesized speech to emphasize specific words, and modifying the pacing of the narration to better express the video’s tone,” NVIDIA wrote. 

NVIDIA is distributing a few of this analysis — optimized to run effectively on NVIDIA GPUs, in fact — to anybody who needs to strive it through open supply by means of the NVIDIA NeMo Python toolkit for GPU-accelerated conversational AI, out there on the corporate’s NGC hub of containers and different software program. 

“Several of the models are trained with tens of thousands of hours of audio data on NVIDIA DGX systems. Developers can fine tune any model for their use cases, speeding up training using mixed-precision computing on NVIDIA Tensor Core GPUs,” the corporate wrote.