Apple AI experts work to solve big speech-translation problems


Translation Unsplash
Machine translation isn't easy. But Apple's working to solve it.
Photo: Romain Vignes/Unsplash

Apple is seemingly looking to step up its work in speech translation, showing one more way that Cupertino is looking to incorporate cutting-edge approaches to artificial intelligence into what it does.

In a recent paper, published to the pre-print server arXiv, two Apple research scientists describe the challenges in the field, carrying out a fundamental survey of the problems with current speech translation.

No, they don’t have a magical solution. But the fact that Apple gave them such a task suggests the company takes seriously the need to hone its tools in this area.

The paper, titled “Speech Translation and the End-to-End Promise: Taking Stock of Where We Are,” was published April 17, 2020. Its two authors, Matthias Sperber and Matthias Paulik, work for Apple as, respectively, a Siri machine translation R&D scientist in Germany and a senior manager in Cupertino. Together, their skill sets cover automatic speech recognition, machine translation, linguistic annotation, speech-to-text technology and more.

They are therefore well-placed to take stock of the current state of the art when it comes to machine translation. In their paper, they note that recent “end-to-end modeling techniques” have improved AI machine-translation systems. But they also face problems, “due to compromises made to address data scarcity.” They make several technical suggestions for ways to improve these speech-translation systems, potentially improving their efficiency.

Unlike other Apple AI papers, this one does not suggest a new approach developed by Apple. Instead, unusually for the secretive Apple, it is an attempt to open up a dialogue about how the field can be moved forward for all involved.

Since Apple doesn’t typically do this, it’s remarkable in itself. But the paper also underlines how Apple wants to improve its own speech-translation capabilities.

Apple is taking speech translation seriously

Over the past several years, Apple has worked hard to increase its AI capabilities. It has hired experts in the field such as Carnegie Mellon researcher Russ Salakhutdinov in 2016. In 2019, Apple hired Google AI expert Ian Goodfellow as its new director of machine learning. Recently, Apple acquired Voysis, an AI startup it hopes will help improve Siri’s natural-language technology.

Meanwhile, Apple launched a quasi-academic blog, the Apple Machine Learning Journal. This gives Apple’s AI researchers the chance to publish some of their research.

Tools like Siri already come with machine translation abilities baked in. However, these abilities are nowhere near as prominent as the work of companies like Google. It’s not clear exactly when Apple is going to be able to show breakthroughs in this domain. But unleashing two of its top researchers to investigate the problem certainly shows that the company is taking it seriously.