About the talk
Speech recognition remains one of the most popular applications of machine learning, but that doesn't mean the technology is well understood by an average business user.
To find out about the latest trends in this rapidly developing field, AI Business sat down with Sam Ringer, machine learning engineer at Speechmatics, a British company which develops automatic speech recognition software based on recurrent neural networks and statistical language modeling.
We talked about unsupervised learning algorithms, the cost of training for transcription systems, and approaches to data labeling at Tesla and Facebook.
Hello, we're here at the Javits Center off of the second day of AI Summit New York and I'm talking to some ringer machine learning engineer at 00:04 speechmatics. I think I previously you talked about biscuits the latest disruptions in speechwriter. Yeah. Yeah, so I 00:12 think I told you that the company one of the trends that we're 00:22 starting to see take offense which industry is that there's been a lot of progress. You know, how long has been a lot of 00:32 motorcycle development in 00:41
any language, 00:44 but the same algorithms to speech so, how can we learn something speech without using a protector? 00:53 is learning algorithms to see the same break through 16 and I'll pee but just to bring those three speech about to turn this is really starting now, I 01:04 think going to see a lot more of it in 2020 and 2021. That's a trend that is not quite mention yet, but it's headed that way we can watch 01:13 one of these is the super expensive 01:23 labeled as 01:33 hard as expensive as you go to make sure you know that covers over at the name see if you're interested in it can be mislabeled. 01:41
The problem is people you saw it and it was that a video of a Tolkien is being transcribed. Some people started coughing Moto's would think 01:51 that the water flows just being said that it's a close up close on his face was labeled someone who said this is an example that 02:01 just by having labeled and she's the possibility that the labels to be wrong, which is a big problem that you know, I think is 02:10 underappreciated is why do humans ready that need very much labor day so we can just go to explore the wild to pick up toxins in a very nice way. So 02:20
it is at least we should do we pick your voice audition cities these problems because we can do it. So I need a we can send you a transcript 02:28 stuff positively regardless of noise in the background and Accents in and stuff like that so I can never talk to you when I think not only will we 02:37 save money because you don't have to come up with her because it's just a smoothie or not and labels better results. I'm quitting. 02:47 They 02:56 labeling data is their special is there right there on the Shelf? What do you think? All of those guys are going to get out of business? 03:07
What are the dates you're labeling companies like 03:17 lots 03:21 of money to get. Data labels United States are really really really really expensive are just walking and losing 03:33 a lot of progress in that sell trading cards box humans can learn to drive in about 20 hours. Okay, so that's why 03:43 we should be able to solve that problem using a way to order. 03:52 So one of us has a big Labor Day, so I'm not sure and I don't expect. 04:00 Traffic lights. What someone said this is the traffic like I think that it's going to be huge amounts of body when I have every company like Facebook 04:17
or the other and nothing has a very specific source of data. You know, how about you at about me? And it can Jordan size that humans wouldn't be able 04:25 to do otherwise so it can be labeled quickly by humans, I think about building businesses just to solve that 04:35 problem at a time. So, 04:45 I think it's a 04:55 set-up 04:55 you don't like years more than you can shake a stick at right now. Happened because there is so 05:06 much text. I know we can see strong hours in United Vision and speech that label out 05:15
in times about an issue just need less high quality Labor Day. She's got 05:25 started. It's just easier to climb a big pool a beta tester. He would like logic companies like, you know, Google maybe Tessa. 05:35 I think I was really counting method of interacting with the 05:44 cology because it makes sense in it. So you say we interact with each other by speaking to each other. We don't need to read each other like he was 05:54 going to have a conversation of a text to each other so, you know appeared in time. Like I really like we should have just like the late 1960s. In 06:02
America. Folders in a casket open a mouse in the oldest 06:11 of what I did yesterday we can 06:21 interact with our computers any more Notch Road by voice thing. I can be really powerful and you can you can salsa stay in things like Alexa if we can 06:30 I don't think it's going to quite replace the heart attack affect a bit like a people going to be really whiten code using my voice. Like I didn't 06:39 think so, but the way we talk to each other which is a voice. I think that's really changed. Our 06:49
relationship was technology, you know, you know, I'm a mouse. Can you tell me 06:59 did not see the keyboard? So it'd be cool to think that they could be a similar sort of Revolution the rising, you know. Higher voice read realizes 07:08 its full potential how 07:18 to solve the problems but the environments 07:28 so that we can do that. I can come to the transcription and the understanding pop sold in 07:36 those things have been started getting so much better than Alexa 07:46 things like the rest of the 07:55
world. What are the benefits and 07:58 drawbacks of being established technology hub? I'm passing I really enjoy it. So 08:08 it is quite harsh 08:14 say if you're in the city. Cambridge amount of these people knocking around. It just makes stalking a business easier University 08:24 town is the University that load of people walking speechmatics was at the University. I found it was a ton of the University so really strong 08:34 ties that just makes it easier. Nassar pipeline from ringside straight inside the kibosh that thing is out and it's much stronger through being 08:44
somewhat like Cambridge. I don't think it's just living there is a yard walls. So they're thinking about 08:52 things and she's not a people who care about but being a community where there are other people you care about yourself is it? 09:01 Fox I 09:11 yeah, 09:12 I think that can be problematic benefits of being in the street in town is typically students instead of more like unrealized potential 09:25 energy and you can serve them is that coming out of University before you thought that makes it easy healthy ecosystem and Sons if you're 09:35
walking and Cambridge in you can say Wallah Costco job offers from XY and Z they pay them there. So I got to walk like this. Just remember I think it 09:44 gives you more which is probably a good thing that you should always think they're going to 09:54 come out ahead in that competition. 10:04 Daddy. 10:05 People are going to leave you fast. So maybe some of those are the facts. 10:16 She's 10:26 not even just makes commissioning vitamin C and Olivia and we use a lot of container is technology. So being able to 10:39
stay for this. If you're still talk about makes life a lot easier instead of having to use an exacto Windows system on something that 10:48 affects me more like I'm using on the same pytorch evangelist may just 10:58 doing research something having an open sore and very active community in the phone with pytorch just let you 11:07 do custom stuff with a lot more flexibility than if you were trying to use them for the proprietary solution 7 Machine learning 11:17 on the research site to share your research. So, you know, you can go to like maybe you're about to attack and you want to read the paper you might if 11:27
you're not affiliated with University of a hundreds of pounds. Paper, which is kind of insane right in the machine Learning Community where everyone 11:35 for free or not easy to learn stuff, right? And you know, I I mean that interested in some area I can just download 11:44 the funny papers and read it wasn't because that's the sort of community and its expectation that a shower research because of that and we 11:53 can rely on each other. So I think that's a real benefit a real underappreciated reason as to why machine. He's moving at the pace it is because the 12:03
resources available all the time anyone can look at it for free and it's only moves. The people who can walk outside, you know, the universe is maybe 12:12 from Cambridge University anyone in the world and if it were these papers, so I posted post, what has been the most unusual 12:21 there like machine learning projects. You have I ever worked on 12:31 the case wide one ought to be out. There 12:34 should be capitalized it should have, is?. So we are using some of the latest techniques and I'll text you then. I'll Pee Wee system 12:43
which is the loudest song, you know, I got quite upset with punctuation for 6 months and debating whether I like she did and then I'll be at 12:53 my 6-week we have to like me but as I see different languages 13:00 support the punctuation to Korean and you know, I know the Korean speak and read in Korea. No, I might confuse to 13:09 tell me what I would like is Harry Potter. And should that be a. Bad? Maybe I should be a? I know this is kind of a surreal moment how somebody got 13:19
some Korean and I like to meet Ron at the place that you are punctuation system in Korean. I don't like the next day. I was trying to lose back and 13:26 trying to figure out for the stuff out and then I was rushing the day off to that require highly leveraged. Hey, I can find 13:33 some of these problems without maybe having a huge contracts with 13:43 what 13:45 time is it supposed to spot patterns in the Korean things to do in like a patient. Is that symbol? 14:05
Buy this talk
Buy this video
With ConferenceCast.tv, you get access to our library of the world's best conference talks.