Finally, No More Need to Learn English?

Great news recently for business people who've struggled to learn English: Facebook has pulled off fluent translation between English and Hokkien. They call it "Hokkien," but given that the engineers who built it are Taiwanese, you might as well read it as a Taiwanese (Taigi) translator.

Facebook's metaverse products have taken a beating in ridicule and skepticism, but this translation system is a reminder that the company can still develop on multiple fronts at once — you underestimate a tech giant's underlying capability at your own risk.

What's So Impressive About This? Aren't There Already Piles of Translation Gadgets on Crowdfunding Sites?

Facebook's breakthrough comes down to one basic fact: most general-purpose translation systems have always had to route through English as a hub, translating the source language into English first, then from English into the target language. Plenty of people have tried building systems that translate directly between two non-English languages, but the results have always been worse than going through English. It's the same as literary translation in the physical world — works tend to move between major languages first, and only then trickle down into smaller ones.

Facebook claims its multilingual model actually outperforms the bilingual approach. The upside is huge: overall system performance improves substantially, and smaller languages that were previously very hard to tackle start getting noticeably better translation quality. Many of the world's languages have no standardized writing system at all, and by showing a working direct English-Taiwanese translation, Facebook is essentially inviting you to imagine a future where the world's more than 3,000 spoken languages can eventually all be translated directly, without an English detour.

So Do We Still Need to Learn English?

This isn't the first time we've seen an AI breakthrough get hyped to the sky and then quietly stall out. Tesla's Elon Musk has repeatedly declared that "this year" will be the dawn of full self-driving — and yet from 2014 to today, that dawn still hasn't arrived; he's even joked about it himself in interviews. That's because the last mile of full artificial intelligence always turns out to hide far more unknowns than it looks like from the outside — it seems fine right up until it suddenly hits a ceiling.

Speaking Taiwanese Works Too — Facebook's Big Translation Breakthrough

Just how hard is voice AI, really? Look at the seemingly all-powerful Amazon. Amazon's data holdings rank among the largest in the world, and yet when it first set out to build Alexa, the company ran into a shocking realization: it simply didn't have enough voice data. Amazon running short of data sounds about as absurd as America running short of corn — but a shortage is a shortage. Nearly everything Amazon had on hand was transactional data, because, well, nobody talks out loud while reading an e-book or browsing an online store.

On top of that, Amazon founder Jeff Bezos wanted the Echo speaker to work reliably in the actual, messy environment of a real home — a very different challenge from what Google and Apple were solving for, since we speak directly into a phone's microphone when we use Siri. The average household is noisy and voices come from far away, so correctly triggering a voice assistant under those conditions was borderline black magic using 2014-era technology. Smart speakers demand extreme accuracy: get the wake word wrong and the whole thing simply doesn't activate, and if two out of every ten sentences go unrecognized, the experience becomes unusable.

What followed was a brutal grind of throwing money at the problem — the Alexa team eventually swelled to roughly 2,000 people, and to keep Google and Apple from catching wind of the project, Amazon layered on extremely tight secrecy measures that cost even more. Early development went badly. When Bezos reviewed the team's projected timeline and realized it could take as long as 20 years to finish, he was stunned, and reportedly tore into the team over it.

Building the Smart Speaker — Amazon Grinding It Out With Brute-Force Spending

In the end, it was Amazon's classic playbook — spend enormous sums to get big things done — that solved the problem. They acquired smaller companies, including Poland's Ivona for $30 million, and spent heavily worldwide buying up voice data. Even that wasn't enough, so they eventually rented houses, filled them with microphones and the kind of ambient noise-making appliances you'd find in a real home, and hired people to go inside and simulate everyday life, reading scripted lines out loud for the machine to learn from — essentially a reality-TV concept for voice recording. Because they needed a random cross-section of accents and age groups, the constant foot traffic in and out of these houses got so heavy that neighbors suspected a drug operation and called the police. Under the strict secrecy rules, Amazon's own employees couldn't even explain to the officers what they were actually doing.

All that effort paid off — Alexa became a huge success. By 2018, the number of conversations users had with Alexa had grown by tens of billions, and the more people talked to it, the more accurately it learned, boosting the machine's comprehension by 20%. Today Alexa is a completely unremarkable fixture of smart home devices; hardly anyone bothers filming themselves chatting with it anymore, because it's simply become normal. Smart speakers are now just part of daily life. All that Herculean effort was spent purely to get a speaker to understand you when you talk to it.

So back to the original question: when do we actually get to stop learning English? Right now, Facebook's demo of Taiwanese-English translation has just crossed the threshold of being genuinely usable. If you already know a bit of simple English, within a year or two every video-conferencing system should offer live translated captions between major languages — imprecise, but workable, and only workable. Fully seamless real-time conversation is probably seven or eight years out, and for anyone with exacting standards, it'll take considerably longer than that.

As the self-driving saga shows us, AI's last mile always has more problems to iron out than expected. So "good enough for casual everyday use" should arrive fairly soon, and skipping foreign-language learning entirely is plausible for babies being born today. But for those of us who are already adults, going completely without English within the next five years just isn't realistic. So for the time being, we still need to learn English. Rest easy, A-Ti and Teacher Hsu Wei (阿滴, 徐薇) — Taiwan's beloved English tutors — your jobs are safe!