[scottlocklin]

Please characterize the amazing improvements the last 20 years have brought to these fields.

Your analogy is faulty: Haber Bosch was obviously something very different and new. Nothing in use today differs appreciably in accuracy or capability in speech to text or machine translation as it was 20 years ago.

[virgilp]

Care to tell me one single speech to text product that had at least a horrible 50% accuracy for Romanian 20 years ago? Because today I can dictate my messages to Google - that, plus the occasional correction, is faster than typing.

[virgilp]

No, it did not work adequately, not even close, not even in English. Maybe if you gave it good conditions (god forbid any background noise!) and spoke clearly you could get some decent results. But that's not "adequate". How many people did you know using speech to text 20 years ago?

It's exactly the difference you suggested with Haber Bosch vs precursors - "we know how to do it" is a far cry from "it's cheap & repeatable enough to be actually available in practice, to everyone".

[scottlocklin]

Lol, OK dude. Rabiner is still a classic and useful textbook on this topic, and it was published in 1978.

[virgilp]

So are all the stats books. What exactly is that supposed to prove? Basic things seldom change, and remain useful.

[scottlocklin]

I guess I'm always shocked and amazed how hypnotized people are by their ipotatoes, and how little technological perspective people actually have.

I was there back in 1990s using speech recognition software which worked just like it does today, with vastly more than 50% accuracy. Dunno where you were. Romania I guess.

[virgilp]

Yes. Romania. What you think software didn't run in Romania?

What did you use, Dragon Speech? Maybe it worked better for you? I can assure you for non-native speakers it was pure crap. Also, what was the usage in western world? I know they had _some_ sales, but I don't remember it exactly taking off.

Lastly - you're obviously willfully ignoring progress, like translation for instance works so much better with RNNs and attention. The basics of RNN were indeed there (LSTM was just published - though e.g. GRU is more recent), but attention was not. And attention is critical. That's just one of the things that's not "basically the same" and "worked just as well 20 years ago".

[scottlocklin]

My apologies for not knowing the details of Romanian speech to text gizmos, but for English, Spanish and other more popular languages speech to text worked adequately, using contemporary hardware even back in the 1980s when Romania was still part of the Warsaw Pact. The fact that google or whatever has gotten around to training a vector quantizer on Romanian now isn't really a sign of actual technological progress. It's just last mile.