Google voice recognition: reduced by 30% transcription errors
Voice recognition is one of the projects supported by Google that has achieved, in recent years, the implementation of the system in a variety of fields with an ever-smaller percentage of error.
From the “Gboard“, virtual keyboard for Android and iOS, or from the same homepage of Google until you get to the tangible products as the speaker is intelligent, “Home“, the challenge of the Mountain View giant continues and is enriched by ever more unimaginable.
As stated by Jeff Dean, Senior Fellow of Google, the voice recognition is based on the development of “neural networks“, systems capable of learning large amounts of data from users such as, for example, words or fragments of speech and use them to make inferences, just as is happening in the mind of a child or of a person during the learning of a language.
Today, the continuous training of the neural networks has led to a reduction in the error rate for the word of about 30% compared to 2012, the year in which they are launched the first time with Android Jelly Bean.
With regard to the competition, Google in 2015, reported an error rate per word of 8% is located in front of the other two giants like Microsoft and Apple who still hold to, respectively, 6.3% and 2%. It is not possible of course to make a real comparison among the three because data derived from studies targeted different with different objectives.