Google artificial intelligence division decided last year to join forces with NOAA (U.S. National Oceanic and Atmospheric Administration) to use artificial intelligence to better understand the song motifs of humpback whales swimming in Hawaiian waters. Now the Google team is working on a new project to help users with language problems.
Artificial intelligence, Google makes speaking easy
Incorporating the importance of virtual assistants and speech recognition services, they analyze how they can use these dynamics to improve the communication and interaction of these users.
As Google explains in an explanation in its Artificial Intelligence blog, Parrotron is a sequence-based, attention-based model trained in two phases using two parallel sets of input/output pairs.
So far, people with speech problems have been excluded from these technologies because the AI has not been trained to look at different scenarios. To solve this problem, Google developed Parrotron:
Parrotron, an end-to-end trained speech-to-speech conversion model that maps an input spectrogram directly to another spectrogram without using a discrete intermediate representation. This model can be trained to normalize the speech of any speaker, regardless of accent, prosody, or background noise. Basically, that is an end-to-end artificial intelligence network that transforms atypical speech patterns into synthesized and fluid speech for people with speech disabilities.
How Parrotron works through AI?
The Google team tested this technology with different users, and the Parrotron system reduced the word error rate from 89% to 32%. A more than promising result that shows the potential of this tool.
We can read all the steps of these tests and the technical details of the training in Google’s report.