It works in part like the human brain.
Text-to-speech conversion is becoming increasingly clever, but there’s a problem: it can still take plenty of training time and resources to produce natural-sounding output. Microsoft and Chinese researchers might have a more effective way. They’ve crafted a text-to-speech AI that can generate realistic speech using just 200 voice samples (about 20 minutes’ worth) and matching transcriptions.
The system relies in part on Transformers, or deep neural networks that roughly emulate neurons in the brain. Transformers weigh every input and output on the fly like synaptic links, helping them process even lengthy sequences very efficiently — say, a complex sentence. Combine that with a noise-removing encoder component and the AI can do a lot with relatively little.
The results aren’t perfect with a slight robotic sound, but they’re highly accurate with a word intelligibility of 99.84 percent. More importantly, this could make text to speech more accessible. You wouldn’t need to spend much effort to get realistic voices, putting it within reach of small companies and even amateurs. This also bodes well for the future. Researchers hope to train on unmatched data, so it might require even less work to create realistic dialogue.
