Google presents Translatotron, a tool that can translate spoken word content from one language to another without written transcription.
Google Translatotron. Behind this name as sexy as a striptease from R2D2 hides one potentially revolutionary tool.
In a post on the company's artificial intelligence blog, researchers present in detail the progress that Translatotron offers. Simply put, this would be the first tool ever developed translate oral content directly from one language to anotherwithout going through a written transcription.
Usually, traditional translators insist that the speech is first put on paper, in this case on the screen, before it is translated, then spoken orally by a robot. It is also these multiple forced phases that regularly distort the results.
In this case, Translatotron can look like a spectrogram. It produces a visual representation of the audio frequencies. This is done and thanks to machine learning, a series of artificial neurons will generate a new spectrogram whose frequencies correspond to the output language. No text, fewer steps, faster translation.
An imitation translator
As if that wasn't promising, Google wanted to push the vice a little far. AI researchers had the idiotic idea of integrating an encoder that could at least reproduce itself to imitate the speaker's voice. The technology is still in its infancy and the reproduction is fairly approximate, although some intonations, timbre or cadence of the dictation are fairly close to the original model.
If you are curious about the result, you can listen to some examples of translations from Spanish to English on this site.
We don't know yet what Google intends to do with this technology. It is a safe bet that the teams of the Mountain View company will continue and perhaps allow their work Google Translate is taking a huge stepif the results were spectacular.