|
Briefing
|
- WaveNet in Google Assistant – Deployed WaveNet, Deep Mind’s new deep neural network capable of producing more realistic sounding speeches than existing techniques for Google Assistant, first integration of WaveNet in commercial product
- Limited Implementation – Implemented only for U.S. English and Japanese languages
- Improved Model – 1,000 times faster, producing 20 seconds of high quality audio in one second, compared to 0.02 seconds of audio per second in original version
- Natural Sounding Speech – One sample in U.S. English received mean opinion score (MOS), cumulative score received from test subjects asked to rate recordings, of 4.347 out of five, close to rating of human speech at 4.667
|
