The new recognizer designed for children's voices, supports both the company's TrulyHandsfree phrase spotting technology and TrulyNatural large vocabulary continuous speech recognizer. With this technology advancement, says the company, developers of apps, children's toys, kid's wearables, and education technology can implement voice control technology with unparalleled accuracy and privacy due to the company's AI-on-the-edge architecture.
Accurately recognizing children's speech is challenging because it differs from adult speech in many ways. A scarcity of available training data makes this problem even more difficult to solve.
The company says that over many years it collected and analyzed significant amounts of children's speech to better understand and model the specifics of how children talk. Initial testing on a corpus of spontaneous kid's speech, reveals up to a 33% reduction in word error rate, when compared to an adult speech recognition model.
"Sensory has some of most talented technologists in the speech industry," says Todd Mozer, CEO at Sensory. "We challenged the team to create a private and accurate recognizer for kid's speech and they delivered. This opens up new and fun voice enabled products for kids of all ages."
Developers can now access child speech models, as well as the company's adult speech models, within Sensory's VoiceHub developer portal. The flexibility of VoiceHub enables direct export to many supported DSP and microcontroller formats, including newly added ICs from Generalplus Technology, a worldwide supplier of integrated circuits for speech and toys.