Is the era of artificial speech translation upon us?
What used to belong only in science fiction—devices that let people converse across different languages—has finally arrived. The question is how well it actually works.
The biggest obstacle facing artificial speech translation, Ingmar Voss tells me, is noise. In a laboratory or a quiet meeting room a device may pick up speech accurately enough, yet it falters against exactly the sort of background din I can hear swirling around Professor Voss as he talks to me from the main railway station in Taipei. Over a crackly connection I keep losing his English, a reminder that the two of us sit almost 9,000 kilometres apart, and that distance still gets in the way of understanding even when, as now, both people share a language. The future, plainly, has not arrived. Had it done so, Voss could have spoken more easily in his own Norwegian while I heard everything come through in English.
Voss is a professor of computer science at the Bergen Institute of Speech Technology, where he and his colleagues have long delivered lectures in Norwegian that students follow in English through an electronic translator. Because the translator throws the text onto students' laptops or phones to read, what happens is not far from subtitling. Several things help: the lecturers speak clearly, no rival chatter competes for attention, and from one year to the next the content barely changes.
Artificial speech translation is far from a new dream. Back in 1981, the novelist Halvard Renn gave his science-fiction story The Lantern-Bearers of Vega a small creature called the 'murmur-moth', which a listener had only to hold against the ear to understand any tongue in existence. The moth became shorthand for the kind of gadget that enthusiasts imagine decades before it can actually be built, much as people once pictured wristwatches that could take telephone calls: things unimaginable in one generation and unremarkable in the next. Lately, riding a surge of progress in machine translation and voice recognition, the first devices resembling a rough murmur-moth have begun to surface.
For the moment, though, such gadgets are treated as intriguing curiosities rather than moves toward what Voss calls a 'language-transparent society'. Typically they are household items, or tools for hotel check-ins, offering a handy substitute for a traveller's broken English; the social role matters more than how well the translation performs. Professionals, by contrast, 'have far less patience in a conversation', notes Diego Marchetti, founder and chief executive of Lucent Speech. With them in mind, his company is readying a fresh model for professional use, built around sharper speech recognition, more faithful translation and a shorter wait before the translated words emerge.
Holding a conversation requires each speaker to wear an earpiece—Lucent calls its version the 'Relay'. 'People are reluctant to share one of the earphones with someone they don't know,' Marchetti says, an attitude that should hardly come as a shock. The awkwardness would disappear if such earpieces grew common enough that any stranger you met would probably already be wearing a pair. Whether that comes about, and how soon, is likely to hinge less on the earpieces in isolation than on how widespread voice-controlled gadgets and machine translation become overall.
Voss stresses how much certain East Asian countries matter here, noting that in places such as South Korea voice translation has genuinely caught on, backed by a variety of systems. Even so, much ground remains to be covered. A translation system has to run simultaneously—the way an interpreter's voice rides over a foreign politician interviewed on television—rather than in chunks that force each speaker to halt every few sentences until the translation catches up. It must also work offline, for the many moments with no internet, and calm worries about the volume of private speech piling up on distant servers after being sent away for processing.
Beyond the physical hurdles like noise, these systems will have to acquire a degree of social awareness, speaking to people in a manner that suits them. In certain cultures, for instance, a person's academic standing calls for grave respect, and courtesy demands that it be honoured. A translator alert to such etiquette could spare its users from having to learn every one of these varying social conventions. It might, into the bargain, keep local customs alive, holding back the creep of the manners that travel with international English, among them its eagerness to move quickly to first names.
None of this, however, will lead professors and other specialists to hand their sense of language over to a program. Should the technology ripen into something seamless and everywhere at once, it will if anything make linguistic skill more valuable, not less. What it will mean for people's family lives and relationships is harder to predict, although one intriguing possibility is that it might dissolve the barriers of language that so often open up between the generations once a family has migrated, stranding grandparents and grandchildren without a tongue in common.
Whatever it ends up being used for, the technology will never match the genuine article. Voice-morphing may reproduce a speaker's voice, yet the lips will still be out of step, like a clumsily dubbed film. That mismatch only highlights how much a shared language is worth, and how much is gained by learning one. Having a language in common can foster a feeling of belonging and fellowship, as it does for the scientists worldwide who now lean on English as their shared tongue where once their forerunners relied on Latin. The practical case for a single common language may fade, but the social rewards of sharing one will endure. No program, in the end, will replace the fine-grained yet essential understanding that only real command of a language brings.
