According to a new report, Google Research is working on a new, personalized AI assistant that can work offline.
Google published a paper detailing the new speech-recognition system titled Personalized Speech Recognition on Mobile Devices on Friday. The Google Research team was given the task of creating a system that does not rely on a server, works on your device accurately without leaving a major footprint, and is based on voice commands and dictation. The new system takes up 20.3 megabytes and is underpinned by machine-learning techniques. To train it, the researchers used approximately 2,000 hours of anonymized Google voice search traffic, totaling up to 100 million utterances, and added in noise from YouTube to imitate real-life conditions. Google will present its report at the 41st International Conference on Acoustics, Speech and Signal Processing which begins on March 20.
In other Google AI news, the company's DeepMind unit's AlphaGo program has continued to take on human Go champions. AlphaGo is up 3-1 in its current series against 33-year-old South Korean Go champion Lee Se-dol, with Lee giving the AI its first human loss in the fourth of a five-game match in Seoul; last year, the AI beat European champion Fan Hui 5-0. According to DeepMind founder Demis Hassabis, AlphaGo made some mistakes. "#AlphaGo thought it was doing well, but got confused on move 87. We are in trouble now..." tweeted Hassabis, who later clarified that the AI made a mistake on move 79 but only realized it by 87.
Photo: © rvlsoft - 123RF.com