Integrity Languages


Reassessing Speech Translation

By: Jonathan Downie    Date: February 24, 2020

Back in October, I wrote a blog post in which I admitted that I had jumped the gun in my assessment of remote interpreting. Now, after writing a book on speech translation and pointing out the flaws in a recent BBC article on the subject, it is time for me to go through the same process for speech translation. And the results are … slightly more complex.

Two Worlds

In my book, Interpreters vs Machines, I deliberately concentrated on the basic operating principles of all speech translation solutions and on the research that was available at the time. After that, I deliberately chose to focus on the claims of commercial speech translation solutions, since they were getting the most attention by the media and by professional interpreters. It turns out that my decision was actually pretty sound.

Academic research on speech translation is continuing quietly and is making steps to deal with one or two of the major issues I discussed in the book, especially the losses that happen when you turn speech into text, run it through machine translation and then turn the results into speech again. Apart from the Google Translatotron, which is yet to be subject to proper public assessment, there is an upcoming conference that seems to be pushing the idea of going directly from speech to speech.

The commercial world, however, continues to produce some quite remarkable claims. Take this one from a recent video by Waverly Labs:

Doing our research, we studied the tools used by professional interpreters, by taking inspiration and going one step further, we gave ambassador the capability to deliver natural, professional grade translation.

While we could quibble about the meanings of the terms here, the phrase “natural, professional grade translation” is a pretty bold claim. Either it means that they are claiming to have matched the quality of professional interpreters, which their own CEO admitted they haven’t or they are claiming their system is good enough to be used by professionals. In the latter case, one might ask whether those professionals should be persuaded to switch from using professional human interpreters.

In either case, it is clear that the purveyors of commercial speech translation are making incredibly bold claims without citing any empirical evidence. How long it will be before we have a repeat of the embarrassing Tencent incident or even the Microsoft “human parity” blunder is anyone’s guess.

Language Access

But, despite questionable marketing practices, speech translation does have a place. Yet again, it helps to turn to the always thought-provoking, Sarah Hickey, of Nimdzi Insights and now Troublesome Terps. At the #Conf1nt100 conference in Geneva, she pointed out that speech translation is finding niches in places where professionals wouldn’t be used anyway.

If you run a library, speech translation can help you achieve basic communication with patrons from other countries; in emergency situations, it can allow for simple triage until a human can be found, and of course, speech translation is great for tourists and frequent business travellers.

In short, speech translation is providing language access and doing it in places where that access might not have been previously available. That can only be good news.

What does all this mean for human interpreters? You’ll have to buy my book to discover that but I will say that we need to look beyond the unfortunately flawed coverage and crazy claims of speech translation to spend more time thinking through what researchers are managing to achieve. Basically, for now at least, ignore the marketing waffle and trust the engineers.

Want to know more?

As ever, if you are looking for advice as to how to get the best out of interpreting in your business, looking to build an interpreting team for your next event, or if you are looking for a conference interpreter in the UK working between French and English, drop me an email.

Leave a Reply

Your email address will not be published. Required fields are marked *

6 − 4 =