Comment by com2kid | Hacker News Reader

Comment by com2kid | Hacker News Reader

Parent Full thread

> I can't help but think about speech recognition 20 years ago. Many of the hot software packages claimed something like 96% accuracy, and that sounded great on paper. People thought intelligent voice-computer interfaces were just around the corner, yet here we are in 2017, and Siri/Alexa/Cortana are barely becoming usable (but still frustratingly lacking in many situations).

Computer speech recognition now has a greater accuracy than human speech recognition, but the types of use cases for talking to a computer are much less sensitive to error. Short snippets and phrases that are designed to get something done right now.

Compare ordering paper towels with Alexa, to asking a friend for a paper towel when your hands are messy. If your friend mishears you, they can just look over, see your dirty hands, and figure out you probably asked for a paper towel. Alexa has no such benefit, computers are held to a much higher bar.

> People thought intelligent voice-computer interfaces were just around the corner,

I would argue it isn't recognition accuracy that makes complex scenarios hard. I've seen context aware recognition being rolled out for keyboards recently (Android's keyboard does it now days, it'll correct the past word based on what the next word is), and it seems like Android does the same for speech reco, based at least on observed behavior during use, but I'm just guessing.

The real hard part is making computers smart enough to do useful things.

We are a long way off from

"Book tickets at The Altair for between 6pm and 7:30pm next Friday and add the reservation to my calendar and my wife's calendar"

being possible in all but a few contrived scenarios. (Now that said, the above scenario is getting easier and easier if you ignore voice, ML is good at figuring out emails that have schedules in them, and forwarding them to other people is now simple, everything is much better than 5 years ago!)

Seriously?

>Computer speech recognition now has a greater accuracy than human speech recognition, but the types of use cases for talking to a computer are much less sensitive to error. Short snippets and phrases that are designed to get something done right now.

You just said that speech recognition doesn't work in general. In general is exactly where self-driving has to work. If we just needed self-driving trains, we'd have them already*

*We do have them, but strangely there isn't much demand since human train drivers appear to do an ok job most of the time.

> You just said that speech recognition doesn't work in general.

I replied that the accuracy is high. It turns out, in retrospect, that accuracy and usefulness are two different things.

In regards to automated driving, achieving better than human isn't all that hard. Heck backup cameras with little "warning zone" lines on them are better than human for the one particular task of backing up. Cruise control systems that maintain distance are better than humans. We are incrementally getting there. The progress is much different than with voice reco.

The problem with voice reco isn't recognition technology, that works fine, it is with having computers understand what the hell to do with the voice. Contrast that to driving, where the end goal is easy to list - Rule 1: Don't hit anything. Rule 2: Get to the destination. (Rule 1 is the hard part, Rule 2 99.9% solved!)

With Voice Reco, we got the transcription part down, but... now what? In the example I gave up above, knowing how to make a reservation is painful thanks to market fragmentation (not everything is an API), and people generally don't go and tag all of their contacts with their relationship status. I happen to have my wife under "wife" and my mother under "mother", that simple step alone gives me much more natural usage of voice input.

Then when I saw "OK Google, directions to my Mother's house", well, that still doesn't work for a thousand little reasons[1], even though each and every word was correctly transcribed. (I get a nice Google search instead!)

The set of situations that can happen while driving is far smaller than the set of interactions that can happen over voice when users expect a natural interface. Yes driving is really complicated, but it is possible to get a group together and after a day or so, brainstorm everything that could happen on the road out to 2 standard deviations of likelihood.

It might take an hour+ just to list all the ways someone might ask for directions.

[1] Mainly because no one programmed it to understand that particular way of making a request. It is annoying because I can ask it for directions to home, and that works fine, so I just figured directions to "Contact Name's House" would also work. Of course if I lived in NYC I'd expect "Contact Name's Apartment" to work!

> The problem with voice reco isn't recognition technology, that works fine, it is with having computers understand what the hell to do with the voice.

If I type "remind me at 9am to water the plants", it does the right thing 100% of the time.

If I say "OK, Google, remind me at 9am to water the plants"...not so much.

If it just got "remind me at 9am" 100% of the time, I'd be satisfied. The fact that"water the plants" becomes "watering lamps" doesn't matter--I can figure out what I meant.

But it doesn't transcribe "remind me at 9am" correctly. It knows what to do with that text, but it doesn't get that far reliably enough.

Maybe I haven't tried saying this to humans enough for comparison, but right now, I'm not satisfied with the transcription of even the most common phrase. (Well over 90% of what I say to my phone starts with "OK, Google, remind me ..."

---

For fun, I just tried reading this out to my phone, and it got that phrase all three times:

> If I type remind me at 9 a.m. to water the plants it does the right thing 100% of time if I say OK Google remind me at 9 a.m. to water plants not so much if it just got remind me at 9 a.m. 100% of time I'd be satisfied the fact that water the plants becomes watering clamps doesn't matter I can figure out what I meant but it

Not bad. It cut off partway through, and maybe I need to speak my punctuation, but it's less error prone than normal. Helps to speak very clearly, unsurprisingly.

Isn't that kind of the point though? Yes, computers can hear better than humans, but they fail miserably at understanding the meaning of what they are hearing and responding with the appropriate actions. This is exactly the same issue that is/will be faced with self driving/driverless cars.

Reply on news.ycombinator.com