Google has presented its Parrotron project, a neural network from end to end that transforms atypical speech patterns in a synthesized and fluid language, and that is aimed at people with speech disabilities, according to the company in your corporate blog.
The Parrotron project is focused on speech, and achieves this process without the need to produce text and omitting the step of recognition of language signals (such as the movement of the lips). The objective is that this technology can be used between humans and with automatic language recognition engines (ASR).
This tool is part ofl Euphonia project, which, according to Google, has shown "that voice recognition models can be significantly improved to better transcribe a variety of atypical speech and disartic"Google, as well, has started from virtual assistants and voice recognition services, tools that these people can not use due to their difficulties.
As Google explains in a statement published on its Artificial Intelligence blog, Parrotron has been trained in two phases using two parallel compilations of input / output voice pairs.
For this, the researchers built a voice-to-speech conversion model for standard speech. Later they customized the model, adapting it to the atypical voice patterns of the person with difficulties. They resorted to parallel data derived automatically with a speech-to-text synthesis system (TTS). A text-to-speech (TTS) system converts normal text language into speech; other systems recreate linguistic symbolic representation as phonetic transcriptions in speech. One of the most famous people who have used these systems has been the scientist Stephen Hawking.
Google developed several Parrotron system tests, including one with a Google researcher and mathematician, Dimitri Kanevsky, of Russian origin and deeply deaf to parents with normal hearing, and with Aubrie Lee, an advocate for the inclusion of disabled people who have a muscular dystrophy.
In the case of Dimitri, 15 hours of speech were recorded, which were used to adapt the base model to the specific nuances of his speech. The Parrotron system helped him to be understood by both researchers and the Google ASR system alike. The operation of the Google ASR engine in the Parrotron output significantly reduced the word error rate from 89% to 32%.
Aubrie, on the other hand, contributed 1.5 hours of voice recordings that have been key to exemplify the success of this voice technology.
(tagsToTranslate) tool (t) ia google (t) help (t) save (t) difficulty (t) talk (t) google (t) have (t) present (t) project (t) parrotron (t) application (t) intelligence (t) artificial (t) study (t) pattern (t) person (t) convert (t) language (t) synthesize (t) flow