After the chatbot hangover and the much (too much) wasted time, the time will have to come when artificial intelligence is really going to serve us something that makes sense. Better: it will intervene in the basic questions of our lives, perhaps more important than making the most bizarre requests to a series of linguistic models otherwise prone to error. This is the case of communication and in particular of linguistic translation: Google, for example, is progressing apace in a project called USM, Universal Speech Model, a linguistic model that aims to support a thousand different languages from all over the world.
It is part of a program announced last November, but which now finds important updates. Is called 1000 Languages Initiative and aims precisely to build an AI that supports a thousand languages, “bringing greater inclusion to billions of people in marginalized communities around the world,” explained Jeff Dean, senior vice president for Google Search at the time of the announcement.
CES 2023
From Poland with the translator: this is how the Vasco Translator V4 challenges Google Translate
by our correspondent Emanuele Capone
The program has many notable points. The first is that it would really pay off accessible to a large part of the world‘s population the potential of Mountain View’s artificial intelligence systems and therefore all the infinite tools that it will enrich. The second is that, among other possibilities, it would make over a thousand languages perfectly translatable between them: that linguistic model would also become, in a certain way, a universal translator without any obstacles. And it is no coincidence that among the more than 20 artificial intelligence-enhanced products that will be presented this year at I/O, the developer conference scheduled for May 10, we can expect a first, substantial evolution of Translate. Even to a recent one press conference in ParisBig G confirmed that languages are one of the most interesting fronts for developing new AI-based services.
In short, USM is the maximum level of development of a linguistic model: it was created by including two billion parameters trained on the basis of 12 million hours of speech, i.e. of oral corpus, and 28 billion sentences per hour in 300 languages. And it is already employed on several platforms: Alphabet exploits it on YouTube to automatically generate subtitles and also supports automatic speech recognition with significant results. It is clear that if these tools end up being usable in over a thousand languages, a platform will be created which will itself be a sort of Esperanto, a crucial junction to be integrated into apps and platforms to speak in the world without knowing other languages, with translation systems automatic again more efficient than the current Translate or other similar tools integrated in some widely used apps, such as (just to give an example) in the Vinted chat.
However, some of these languages have been spoken for less than twenty million people. They seem like a lot, but they are few so that current artificial intelligences have enough starting materials available to train properly. For this reason, the development of increasingly advanced and powerful machine learning mechanisms is essential to travel towards a true linguistic inclusiveness of AI. Everything holds.
Also ChatGPTquestioned on the point, confirms that she can “communicate in any language for which there are enough data available in my machine learning database. However, my ability to support a specific language depends on the availability of natural language data and my ability to learn and understand the grammatical and lexical structure of the language.”
The guide
How to create images with artificial intelligence (without being an artist)
by Francesco Marino
USM, which, as mentioned, is used on YouTube“can perform automatic speech recognition not only on widely spoken languages such as English and Mandarin but also on resource-limited languages such as Amharic, Cebuano, Assamese, and Azerbaijani to name a few,” reads the a post on the official Google blog dedicated to artificial intelligence. Nell studio Google USM: Scaling Automatic Speech Recognition Beyond 100 Languages, the behemoth demonstrated a key capability of its AI: employing a large unlabeled multilingual data set to pre-train the model encoder and fine-tuning it on a smaller set of labeled data allows data to be recognized underrepresented languages. Additionally, the model training process is effective at adapting to new languages and data. In other words, it is a system that learning from a polyglot set then surprisingly manages to specialize in languages for which you have less material. And to proceed in this way, quickly, towards the ability to learn a language by exploiting the leverage of the others.
The Universal Speech Model currently hovers around 100 languages but is the starting point for one very large machine translation platform powered by artificial intelligence. Meta is also working on a similar project, but with hitherto less effective times and successes.