Google daughter’s DeepMind claims to be a big leap in generating synthesized
The Google department DeepMind, which focuses on the development of self-learning computer systems, claims to be a great leap forward in the production of computer-generated speech.
DeepMind put machine learning techniques in to a so-called ‘text-to-speech (TTS) system to build.
Such systems have been in existence longer, but the quality of human speech can never approach. Traditional spraakgeneratie works by klankfragmenten together to paste as a word or phrase. However, that can sound unnatural.
The system of DeepMind, with the name WaveNet, works by the waveforms of human speech analysis and recognition of patterns. These are then used to written text to convert to sounds.
DeepMind had a hundred testzinnen, which were spoken by real people, WaveNet and older TTS systems, rate by people. It scored WaveNet in American English, a score of 4.21 on a scale of 1 to 5. The best older TTS system scored a 3,86, while the phrases that real people were recorded an average score of 4.55 were.
Also in the Mandarin Chinese made WaveNet is a great leap compared to existing systems, and the computer-generated speech almost just as well-judged as human speech.
It is unclear whether Google has plans to the DeepMind technology to effectively use, for example, the digital assistant Google Now. A source within the company claims against the Financial Times, that that is still unlikely, because there are extremely a lot of computing power is required for the generation of speech with WaveNet.
On the site of DeepMind are a few sentences to listen to, and is also a paper (pdf) from the researchers to read.